Med Image Anal. 2024 Jun 22;97:103246. doi: 10.1016/j.media.2024.103246. Online ahead of print.

ABSTRACT

Accurate instrument segmentation in the endoscopic vision of minimally invasive surgery is challenging due to complex instruments and environments. Deep learning techniques have shown competitive performance in recent years. However, deep learning usually requires a large amount of labeled data to achieve accurate prediction, which poses a significant workload. To alleviate this workload, we propose an active learning-based framework to generate synthetic images for efficient neural network training. In each active learning iteration, a small number of informative unlabeled images are first queried by active learning and manually labeled. Next, synthetic images are generated based on these selected images. The instruments and backgrounds are cropped out and randomly combined with blending and fusion near the boundary. The proposed method leverages the advantage of both active learning and synthetic images. The effectiveness of the proposed method is validated on two sinus surgery datasets and one intraabdominal surgery dataset. The results indicate a considerable performance improvement, especially when the size of the annotated dataset is small. All the code is open-sourced at: https://github.com/HaonanPeng/active_syn_generator.

PMID:38943835 | DOI:10.1016/j.media.2024.103246

Waste Manag. 2024 Jun 28;186:271-279. doi: 10.1016/j.wasman.2024.06.026. Online ahead of print.

ABSTRACT

Cities are a major source of litter pollution. Determination of the abundance and composition of plastic litter in cities is imperative for effective pollution management, environmental protection, and sustainable urban development. Therefore, here, a multidisciplinary approach to quantify and classify the abundance of litter in urban environments is proposed. In the present study, litter data collection was integrated via the Pirika smartphone application and conducted image analysis based on deep learning. Pirika was launched in May 2018 and, to date, has collected approximately one million images. Visual classification revealed that the most common types of litter were cans, plastic bags, plastic bottles, cigarette butts, cigarette boxes, and sanitary masks, in that order. The top six categories accounted for approximately 80 % of the total, whereas the top three categories accounted for more than 60 % of the total imaged litter. A deep-learning image processing algorithm was developed to automatically identify the top six litter categories. Both precision and recall derived from the model were higher than 75 %, enabling proper litter categorization. The quantity of litter derived from automated image processing was also plotted on a map using location data acquired concurrently with the images by the smartphone application. Conclusively, this study demonstrates that citizen science supported by smartphone applications and deep learning-based image processing can enable the visualization, quantification, and characterization of street litter in cities.

PMID:38943818 | DOI:10.1016/j.wasman.2024.06.026

Ultrasonics. 2024 Jun 26;142:107382. doi: 10.1016/j.ultras.2024.107382. Online ahead of print.

ABSTRACT

Non-destructive testing (NDT) is a technique for inspecting materials and their defects without causing damage to the tested components. Phased array ultrasonic testing (PAUT) has emerged as a hot topic in industrial NDT applications. Currently, the collection of ultrasound data is mostly automated, while the analysis of the data is still predominantly carried out manually. Manual analysis of scan image defects is inefficient and prone to instability, prompting the need for computer-based solutions. Deep learning-based object detection methods have shown promise in addressing such challenges recently. This approach typically demands a substantial amount of high-resolution, well-annotated training data, which is challenging to obtain in NDT. Consequently, it becomes difficult to detect low-resolution images and defects with varying positional sizes. This work proposes improvements based on the state-of-the-art YOLOv8 algorithm to enhance the accuracy and efficiency of defect detection in phased-array ultrasonic testing. The space-to-depth convolution (SPD-Conv) is imported to replace strided convolution, mitigating information loss during convolution operations and improving detection performance on low-resolution images. Additionally, this paper constructs and incorporates the bi-level routing and spatial attention module (BRSA) into the backbone, generating multiscale feature maps with richer details. In the neck section, the original structure is replaced by the asymptotic feature pyramid network (AFPN) to reduce model parameters and computational complexity. After testing on public datasets, in comparison to YOLOv8 (the baseline), this algorithm achieves high-quality detection of flat bottom holes (FBH) and aluminium blocks on the simulated dataset. More importantly, for the challenging-to-detect defect side-drilled holes (SDH), it achieves F1 scores (weighted average of precision and recall) of 82.50% and intersection over union (IOU) of 65.96%, representing an improvement of 17.56% and 0.43%. On the experimental dataset, the F1 score and IOU for FBH reach 75.68% (an increase of 9.01%) and 83.79%, respectively. Simultaneously, the proposed algorithm demonstrates robust performance in the presence of external noise, while maintaining exceptionally high computational efficiency and inference speed. These experimental results validate the high detection performance of the proposed intelligent defect detection algorithm for ultrasonic images, which contributes to the advancement of the smart industry.

PMID:38943732 | DOI:10.1016/j.ultras.2024.107382

Stat Appl Genet Mol Biol. 2024 Jul 1;23(1). doi: 10.1515/sagmb-2024-0004. eCollection 2024 Jan 1.

ABSTRACT

Understanding a protein's function based solely on its amino acid sequence is a crucial but intricate task in bioinformatics. Traditionally, this challenge has proven difficult. However, recent years have witnessed the rise of deep learning as a powerful tool, achieving significant success in protein function prediction. Their strength lies in their ability to automatically learn informative features from protein sequences, which can then be used to predict the protein's function. This study builds upon these advancements by proposing a novel model: CNN-CBAM+BiGRU. It incorporates a Convolutional Block Attention Module (CBAM) alongside BiGRUs. CBAM acts as a spotlight, guiding the CNN to focus on the most informative parts of the protein data, leading to more accurate feature extraction. BiGRUs, a type of Recurrent Neural Network (RNN), excel at capturing long-range dependencies within the protein sequence, which are essential for accurate function prediction. The proposed model integrates the strengths of both CNN-CBAM and BiGRU. This study's findings, validated through experimentation, showcase the effectiveness of this combined approach. For the human dataset, the suggested method outperforms the CNN-BIGRU+ATT model by +1.0 % for cellular components, +1.1 % for molecular functions, and +0.5 % for biological processes. For the yeast dataset, the suggested method outperforms the CNN-BIGRU+ATT model by +2.4 % for the cellular component, +1.2 % for molecular functions, and +0.6 % for biological processes.

PMID:38943434 | DOI:10.1515/sagmb-2024-0004

Technol Health Care. 2024 Jun 20. doi: 10.3233/THC-240550. Online ahead of print.

ABSTRACT

BACKGROUND: Brain variations are responsible for developmental impairments, including autism spectrum disorder (ASD). EEG signals efficiently detect neurological conditions by revealing crucial information about brain function abnormalities.

OBJECTIVE: This study aims to utilize EEG data collected from both autistic and typically developing children to investigate the potential of a Graph Convolutional Neural Network (GCNN) in predicting ASD based on neurological abnormalities revealed through EEG signals.

METHODS: In this study, EEG data were gathered from eight autistic children and eight typically developing children diagnosed using the Childhood Autism Rating Scale at the Central Institute of Psychiatry, Ranchi. EEG recording was done using a HydroCel GSN with 257 channels, and 71 channels with 10-10 international equivalents were utilized. Electrodes were divided into 12 brain regions. A GCNN was introduced for ASD prediction, preceded by autoregressive and spectral feature extraction.

RESULTS: The anterior-frontal brain region, crucial for cognitive functions like emotion, memory, and social interaction, proved most predictive of ASD, achieving 87.07% accuracy. This underscores the suitability of the GCNN method for EEG-based ASD detection.

CONCLUSION: The detailed dataset collected enhances understanding of the neurological basis of ASD, benefiting healthcare practitioners involved in ASD diagnosis.

PMID:38943414 | DOI:10.3233/THC-240550

Niger J Clin Pract. 2024 Jun 1;27(6):759-765. doi: 10.4103/njcp.njcp_862_23. Epub 2024 Jun 29.

ABSTRACT

OBJECTIVES: This study aims to assess the diagnostic accuracy of an artificial intelligence (AI) system employing deep learning for identifying dental plaque, utilizing a dataset comprising photographs of permanent teeth.

MATERIALS AND METHODS: In this study, photographs of 168 teeth belonging to 20 patients aged between 10 and 15 years, who met our criteria, were included. Intraoral photographs were taken of the patients in two stages, before and after the application of the plaque staining agent. To train the AI system to identify plaque on teeth with dental plaque that is not discolored, plaque and teeth were marked on photos with exposed dental plaque. One hundred forty teeth were used to construct the training group, while 28 teeth were used to create the test group. Another dentist reviewed images of teeth with dental plaque that was not discolored, and the effectiveness of AI in detecting plaque was evaluated using pertinent performance indicators. To compare the AI model and the dentist's evaluation outcomes, the mean intersection over union (IoU) values were evaluated by the Wilcoxon test.

RESULTS: The AI system showed higher performance in our study with a precision of 82% accuracy, 84% sensitivity, 83% F1 score, 87% accuracy, and 89% specificity in plaque detection. The area under the curve (AUC) value was found to be 0.922, and the IoU value was 76%. Subsequently, the dentist's plaque diagnosis performance was also evaluated. The IoU value was 0.71, and the AUC was 0.833. The AI model showed statistically significantly higher performance than the dentist (P < 0.05).

CONCLUSIONS: The AI algorithm that we developed has achieved promising results and demonstrated clinically acceptable performance in detecting dental plaque compared to a dentist.

PMID:38943301 | DOI:10.4103/njcp.njcp_862_23

J Cheminform. 2024 Jun 28;16(1):75. doi: 10.1186/s13321-024-00870-9.

ABSTRACT

Conformal prediction has seen many applications in pharmaceutical science, being able to calibrate outputs of machine learning models and producing valid prediction intervals. We here present the open source software CPSign that is a complete implementation of conformal prediction for cheminformatics modeling. CPSign implements inductive and transductive conformal prediction for classification and regression, and probabilistic prediction with the Venn-ABERS methodology. The main chemical representation is signatures but other types of descriptors are also supported. The main modeling methodology is support vector machines (SVMs), but additional modeling methods are supported via an extension mechanism, e.g. DeepLearning4J models. We also describe features for visualizing results from conformal models including calibration and efficiency plots, as well as features to publish predictive models as REST services. We compare CPSign against other common cheminformatics modeling approaches including random forest, and a directed message-passing neural network. The results show that CPSign produces robust predictive performance with comparative predictive efficiency, with superior runtime and lower hardware requirements compared to neural network based models. CPSign has been used in several studies and is in production-use in multiple organizations. The ability to work directly with chemical input files, perform descriptor calculation and modeling with SVM in the conformal prediction framework, with a single software package having a low footprint and fast execution time makes CPSign a convenient and yet flexible package for training, deploying, and predicting on chemical data. CPSign can be downloaded from GitHub at https://github.com/arosbio/cpsign .Scientific contribution CPSign provides a single software that allows users to perform data preprocessing, modeling and make predictions directly on chemical structures, using conformal and probabilistic prediction. Building and evaluating new models can be achieved at a high abstraction level, without sacrificing flexibility and predictive performance-showcased with a method evaluation against contemporary modeling approaches, where CPSign performs on par with a state-of-the-art deep learning based model.

PMID:38943219 | DOI:10.1186/s13321-024-00870-9

BMC Ophthalmol. 2024 Jun 28;24(1):273. doi: 10.1186/s12886-024-03532-4.

ABSTRACT

BACKGROUND: Glaucoma is a worldwide eye disease that can cause irreversible vision loss. Early detection of glaucoma is important to reduce vision loss, and retinal fundus image examination is one of the most commonly used solutions for glaucoma diagnosis due to its low cost. Clinically, the cup-disc ratio of fundus images is an important indicator for glaucoma diagnosis. In recent years, there have been an increasing number of algorithms for segmentation and recognition of the optic disc (OD) and optic cup (OC), but these algorithms generally have poor universality, segmentation performance, and segmentation accuracy.

METHODS: By improving the YOLOv8 algorithm for segmentation of OD and OC. Firstly, a set of algorithms was designed to adapt the REFUGE dataset's result images to the input format of the YOLOv8 algorithm. Secondly, in order to improve segmentation performance, the network structure of YOLOv8 was improved, including adding a ROI (Region of Interest) module, modifying the bounding box regression loss function from CIOU to Focal-EIoU. Finally, by training and testing the REFUGE dataset, the improved YOLOv8 algorithm was evaluated.

RESULTS: The experimental results show that the improved YOLOv8 algorithm achieves good segmentation performance on the REFUGE dataset. In the OD and OC segmentation tests, the F1 score is 0.999.

CONCLUSIONS: We improved the YOLOv8 algorithm and applied the improved model to the segmentation task of OD and OC in fundus images. The results show that our improved model is far superior to the mainstream U-Net model in terms of training speed, segmentation performance, and segmentation accuracy.

PMID:38943095 | DOI:10.1186/s12886-024-03532-4

Sci Rep. 2024 Jun 28;14(1):14965. doi: 10.1038/s41598-024-65750-6.

ABSTRACT

To address the urgent need for agricultural intelligence in the face of increasing agricultural output and a shortage of personnel, this paper proposes a high precision object detection network for automated pear picking tasks. The current object detection method using deep learning does not fully consider the redundant background information of the pear detection scene and the mutual occlusion characteristics of multiple pears, so that the detection accuracy is low and cannot meet the needs of complex automated pear picking detection tasks. The proposed, High-level deformation-perception Network with multi-object search NMS(HDMNet), is based on YOLOv8 and utilizes a high-level Semantic focused attention mechanism module to eliminate irrelevant background information and a deformation-perception feature pyramid network to improve accuracy of long-distance and small scale fruit. A multi-object search non-maximum suppression is also proposed to choose the anchor frame in a combined search method suitable for multiple pears. The experimental results show that the HDMNet parameter amount is as low as 12.9 M, the GFLOPs is 41.1, the mAP is 75.7%, the mAP50 reaches 93.6%, the mAP75 reaches 70.2%, and the FPS reaches 73.0. Compared with other SOTA object detection methods, it has the transcend of real-time detection, low parameter amount, low calculation amount, high precision, and accurate positioning.

PMID:38942940 | DOI:10.1038/s41598-024-65750-6

J Imaging Inform Med. 2024 Jun 28. doi: 10.1007/s10278-024-01184-w. Online ahead of print.

ABSTRACT

The aim of this study was to investigate the effect of iterative motion correction (IMC) on reducing artifacts in brain magnetic resonance imaging (MRI) with deep learning reconstruction (DLR). The study included 10 volunteers (between September 2023 and December 2023) and 30 patients (between June 2022 and July 2022) for quantitative and qualitative analyses, respectively. Volunteers were instructed to remain still during the first MRI with fluid-attenuated inversion recovery sequence (FLAIR) and to move during the second scan. IMCoff DLR images were reconstructed from the raw data of the former acquisition; IMCon and IMCoff DLR images were reconstructed from the latter acquisition. After registration of the motion images, the structural similarity index measure (SSIM) was calculated using motionless images as reference. For qualitative analyses, IMCon and IMCoff FLAIR DLR images of the patients were reconstructed and evaluated by three blinded readers in terms of motion artifacts, noise, and overall quality. SSIM for IMCon images was 0.952, higher than that for IMCoff images (0.949) (p < 0.001). In qualitative analyses, although noise in IMCon images was rated as increased by two of the three readers (both p < 0.001), all readers agreed that motion artifacts and overall quality were significantly better in IMCon images than in IMCoff images (all p < 0.001). In conclusion, IMC reduced motion artifacts in brain FLAIR DLR images while maintaining similarity to motionless images.

PMID:38942939 | DOI:10.1007/s10278-024-01184-w

Sci Rep. 2024 Jun 28;14(1):14903. doi: 10.1038/s41598-024-65322-8.

ABSTRACT

Remote sensing has been increasingly used in precision agriculture. Buoyed by the developments in the miniaturization of sensors and platforms, contemporary remote sensing offers data at resolutions finer enough to respond to within-farm variations. LiDAR point cloud, offers features amenable to modelling structural parameters of crops. Early prediction of crop growth parameters helps farmers and other stakeholders dynamically manage farming activities. The objective of this work is the development and application of a deep learning framework to predict plant-level crop height and crown area at different growth stages for vegetable crops. LiDAR point clouds were acquired using a terrestrial laser scanner on five dates during the growth cycles of tomato, eggplant and cabbage on the experimental research farms of the University of Agricultural Sciences, Bengaluru, India. We implemented a hybrid deep learning framework combining distinct features of long-term short memory (LSTM) and Gated Recurrent Unit (GRU) for the predictions of plant height and crown area. The predictions are validated with reference ground truth measurements. These predictions were validated against ground truth measurements. The findings demonstrate that plant-level structural parameters can be predicted well ahead of crop growth stages with around 80% accuracy. Notably, the LSTM and the GRU models exhibited limitations in capturing variations in structural parameters. Conversely, the hybrid model offered significantly improved predictions, particularly for crown area, with error rates for height prediction ranging from 5 to 12%, with deviations exhibiting a more balanced distribution between overestimation and underestimation This approach effectively captured the inherent temporal growth pattern of the crops, highlighting the potential of deep learning for precision agriculture applications. However, the prediction quality is relatively low at the advanced growth stage, closer to the harvest. In contrast, the prediction quality is stable across the three different crops. The results indicate the presence of a robust relationship between the features of the LiDAR point cloud and the auto-feature map of the deep learning methods adapted for plant-level crop structural characterization. This approach effectively captured the inherent temporal growth pattern of the crops, highlighting the potential of deep learning for precision agriculture applications.

PMID:38942825 | DOI:10.1038/s41598-024-65322-8

J Am Med Inform Assoc. 2024 Jun 28:ocae165. doi: 10.1093/jamia/ocae165. Online ahead of print.

ABSTRACT

OBJECTIVE: Artificial intelligence (AI) models trained using medical images for clinical tasks often exhibit bias in the form of subgroup performance disparities. However, since not all sources of bias in real-world medical imaging data are easily identifiable, it is challenging to comprehensively assess their impacts. In this article, we introduce an analysis framework for systematically and objectively investigating the impact of biases in medical images on AI models.

MATERIALS AND METHODS: Our framework utilizes synthetic neuroimages with known disease effects and sources of bias. We evaluated the impact of bias effects and the efficacy of 3 bias mitigation strategies in counterfactual data scenarios on a convolutional neural network (CNN) classifier.

RESULTS: The analysis revealed that training a CNN model on the datasets containing bias effects resulted in expected subgroup performance disparities. Moreover, reweighing was the most successful bias mitigation strategy for this setup. Finally, we demonstrated that explainable AI methods can aid in investigating the manifestation of bias in the model using this framework.

DISCUSSION: The value of this framework is showcased in our findings on the impact of bias scenarios and efficacy of bias mitigation in a deep learning model pipeline. This systematic analysis can be easily expanded to conduct further controlled in silico trials in other investigations of bias in medical imaging AI.

CONCLUSION: Our novel methodology for objectively studying bias in medical imaging AI can help support the development of clinical decision-support tools that are robust and responsible.

PMID:38942737 | DOI:10.1093/jamia/ocae165

Sci Bull (Beijing). 2024 Jun 12:S2095-9273(24)00407-9. doi: 10.1016/j.scib.2024.06.011. Online ahead of print.

ABSTRACT

Realizing large materials models has emerged as a critical endeavor for materials research in the new era of artificial intelligence, but how to achieve this fantastic and challenging objective remains elusive. Here, we propose a feasible pathway to address this paramount pursuit by developing universal materials models of deep-learning density functional theory Hamiltonian (DeepH), enabling computational modeling of the complicated structure-property relationship of materials in general. By constructing a large materials database and substantially improving the DeepH method, we obtain a universal materials model of DeepH capable of handling diverse elemental compositions and material structures, achieving remarkable accuracy in predicting material properties. We further showcase a promising application of fine-tuning universal materials models for enhancing specific materials models. This work not only demonstrates the concept of DeepH's universal materials model but also lays the groundwork for developing large materials models, opening up significant opportunities for advancing artificial intelligence-driven materials discovery.

PMID:38942699 | DOI:10.1016/j.scib.2024.06.011

Sci Rep. 2024 Jun 28;14(1):14898. doi: 10.1038/s41598-024-65996-0.

ABSTRACT

In this study, in order to characterize the buried object via deep-learning-based surrogate modeling approach, 3-D full-wave electromagnetic simulations of a GPR model have been used. The task is to independently predict characteristic parameters of a buried object of diverse radii allocated at different positions (depth and lateral position) in various dispersive subsurface media. This study has analyzed variable data structures (raw B-scans, extracted features, consecutive A-scans) with respect to computational cost and accuracy of surrogates. The usage of raw B-scan data and the applications for processing steps on B-scan profiles in the context of object characterization incur high computational cost so it can be a challenging issue. The proposed surrogate model referred to as the deep regression network (DRN) is utilized for time frequency spectrogram (TFS) of consecutive A-scans. DRN is developed with the main aim being computationally efficient (about 13 times acceleration) compared to conventional network models using B-scan images (2D data). DRN with TFS is favorably benchmarked to the state-of-the-art regression techniques. The experimental results obtained for the proposed model and second-best model, CNN-1D show mean absolute and relative error rates of 3.6 mm, 11.8 mm and 4.7%, 11.6% respectively. For the sake of supplementary verification under realistic scenarios, it is also applied for scenarios involving noisy data. Furthermore, the proposed surrogate modeling approach is validated using measurement data, which is indicative of suitability of the approach to handle physical measurements as data sources.

PMID:38942986 | DOI:10.1038/s41598-024-65996-0

Sci Rep. 2024 Jun 28;14(1):14917. doi: 10.1038/s41598-024-65703-z.

ABSTRACT

In tuberculosis (TB), chest radiography (CXR) patterns are highly variable, mimicking pneumonia and many other diseases. This study aims to evaluate the efficacy of Google teachable machine, a deep neural network-based image classification tool, to develop algorithm for predicting TB probability of CXRs. The training dataset included 348 TB CXRs and 3806 normal CXRs for training TB detection. We also collected 1150 abnormal CXRs and 627 normal CXRs for training abnormality detection. For external validation, we collected 250 CXRs from our hospital. We also compared the accuracy of the algorithm to five pulmonologists and radiological reports. In external validation, the AI algorithm showed areas under the curve (AUC) of 0.951 and 0.975 in validation dataset 1 and 2. The accuracy of the pulmonologists on validation dataset 2 showed AUC range of 0.936-0.995. When abnormal CXRs other than TB were added, AUC decreased in both human readers (0.843-0.888) and AI algorithm (0.828). When combine human readers with AI algorithm, the AUC further increased to 0.862-0.885. The TB CXR AI algorithm developed by using Google teachable machine in this study is effective, with the accuracy close to experienced clinical physicians, and may be helpful for detecting tuberculosis by CXR.

PMID:38942819 | DOI:10.1038/s41598-024-65703-z

Sci Rep. 2024 Jun 28;14(1):14951. doi: 10.1038/s41598-024-65354-0.

ABSTRACT

Prostate cancer is one of the most common and fatal diseases among men, and its early diagnosis can have a significant impact on the treatment process and prevent mortality. Since it does not have apparent clinical symptoms in the early stages, it is difficult to diagnose. In addition, the disagreement of experts in the analysis of magnetic resonance images is also a significant challenge. In recent years, various research has shown that deep learning, especially convolutional neural networks, has appeared successfully in machine vision (especially in medical image analysis). In this research, a deep learning approach was used on multi-parameter magnetic resonance images, and the synergistic effect of clinical and pathological data on the accuracy of the model was investigated. The data were collected from Trita Hospital in Tehran, which included 343 patients (data augmentation and learning transfer methods were used during the process). In the designed model, four different types of images are analyzed with four separate ResNet50 deep convolutional networks, and their extracted features are transferred to a fully connected neural network and combined with clinical and pathological features. In the model without clinical and pathological data, the maximum accuracy reached 88%, but by adding these data, the accuracy increased to 96%, which shows the significant impact of clinical and pathological data on the accuracy of diagnosis.

PMID:38942817 | DOI:10.1038/s41598-024-65354-0

NPJ Breast Cancer. 2024 Jun 28;10(1):52. doi: 10.1038/s41523-024-00663-1.

ABSTRACT

Tumor-Infiltrating Lymphocytes (TILs) have strong prognostic and predictive value in breast cancer, but their visual assessment is subjective. To improve reproducibility, the International Immuno-oncology Working Group recently released recommendations for the computational assessment of TILs that build on visual scoring guidelines. However, existing resources do not adequately address these recommendations due to the lack of annotation datasets that enable joint, panoptic segmentation of tissue regions and cells. Moreover, existing deep-learning methods focus entirely on either tissue segmentation or cell nuclei detection, which complicates the process of TILs assessment by necessitating the use of multiple models and reconciling inconsistent predictions. We introduce PanopTILs, a region and cell-level annotation dataset containing 814,886 nuclei from 151 patients, openly accessible at: sites.google.com/view/panoptils . Using PanopTILs we developed MuTILs, a neural network optimized for assessing TILs in accordance with clinical recommendations. MuTILs is a concept bottleneck model designed to be interpretable and to encourage sensible predictions at multiple resolutions. Using a rigorous internal-external cross-validation procedure, MuTILs achieves an AUROC of 0.93 for lymphocyte detection and a DICE coefficient of 0.81 for tumor-associated stroma segmentation. Our computational score closely matched visual scores from 2 pathologists (Spearman R = 0.58-0.61, p < 0.001). Moreover, computational TILs scores had a higher prognostic value than visual scores, independent of TNM stage and patient age. In conclusion, we introduce a comprehensive open data resource and a modeling approach for detailed mapping of the breast tumor microenvironment.

PMID:38942745 | DOI:10.1038/s41523-024-00663-1

Brief Bioinform. 2024 May 23;25(4):bbae225. doi: 10.1093/bib/bbae225.

ABSTRACT

Accurate understanding of the biological functions of enzymes is vital for various tasks in both pathologies and industrial biotechnology. However, the existing methods are usually not fast enough and lack explanations on the prediction results, which severely limits their real-world applications. Following our previous work, DEEPre, we propose a new interpretable and fast version (ifDEEPre) by designing novel self-guided attention and incorporating biological knowledge learned via large protein language models to accurately predict the commission numbers of enzymes and confirm their functions. Novel self-guided attention is designed to optimize the unique contributions of representations, automatically detecting key protein motifs to provide meaningful interpretations. Representations learned from raw protein sequences are strictly screened to improve the running speed of the framework, 50 times faster than DEEPre while requiring 12.89 times smaller storage space. Large language modules are incorporated to learn physical properties from hundreds of millions of proteins, extending biological knowledge of the whole network. Extensive experiments indicate that ifDEEPre outperforms all the current methods, achieving more than 14.22% larger F1-score on the NEW dataset. Furthermore, the trained ifDEEPre models accurately capture multi-level protein biological patterns and infer evolutionary trends of enzymes by taking only raw sequences without label information. Meanwhile, ifDEEPre predicts the evolutionary relationships between different yeast sub-species, which are highly consistent with the ground truth. Case studies indicate that ifDEEPre can detect key amino acid motifs, which have important implications for designing novel enzymes. A web server running ifDEEPre is available at https://proj.cse.cuhk.edu.hk/aihlab/ifdeepre/ to provide convenient services to the public. Meanwhile, ifDEEPre is freely available on GitHub at https://github.com/ml4bio/ifDEEPre/.

PMID:38942594 | DOI:10.1093/bib/bbae225

J Am Coll Radiol. 2024 Jun 26:S1546-1440(24)00535-0. doi: 10.1016/j.jacr.2024.06.011. Online ahead of print.

ABSTRACT

Thyroid nodule evaluation using ultrasound is dependent on radiologist experience, but deep learning (DL) models can improve intra-reader agreements. DL model development for medical imaging with small datasets can be challenging. Transfer learning is a technique used in the development of DL models to improve model performance in data-limited scenarios. Here, we investigate the impact of transfer learning with domain-specific RadImageNet dataset and non-medical ImageNet on the robustness of classifying thyroid nodules into benign and malignant. We retrospectively collected 822 ultrasound images of thyroid nodules of patients who underwent fine needle aspiration in our institute. We split our data and used 101 cases in a test set and 721 cases for cross-validation. A Resnet-18 model was trained to classify thyroid nodules into benign and malignant. Then, we trained the same model architecture with transferred weights from ImageNet and RadImageNet. The model without transfer learning for thyroid nodule classification achieved an AUROC of 0.69. The AUROC of our model after transfer learning with ImageNet pre-trained weights was 0.79. Our model achieved an AUROC of 0.83 from transfer learning of the RadImageNet pre-trained weights. The AUROC from the classification model without transfer learning significantly improved after transfer learning with ImageNet (p-value = 0.03) and RadImageNet transfer learning (p-value <0.01). There was a statistically significant distinction in performance between the model utilizing RadImageNet transfer learning and that employing ImageNet transfer learning (p-value <0.01). We demonstrate the potential of RadImageNet as a domain-specific source for transfer learning in thyroid nodule classification.

PMID:38942163 | DOI:10.1016/j.jacr.2024.06.011

Surv Ophthalmol. 2024 Jun 26:S0039-6257(24)00073-0. doi: 10.1016/j.survophthal.2024.06.005. Online ahead of print.

ABSTRACT

Cystoid macular edema (CME) is a sight-threatening condition often associated with inflammatory and diabetic diseases. Early detection is crucial to prevent irreversible vision loss. Artificial intelligence (AI) has shown promise in automating CME diagnosis through optical coherence tomography (OCT) imaging, but its utility needs critical evaluation. This systematic review assesses the application of AI to diagnosis CME, specifically focusing on disorders like postoperative CME (Irvine Gass syndrome) and retinitis pigmentosa without obvious vasculopathy, using OCT imaging. A comprehensive search was conducted across 6 databases (PubMed, Scopus, Web of Science, Wiley, ScienceDirect, and IEEE) from 2018 to November, 2023. Twenty-three articles met the inclusion criteria and were selected for in-depth analysis. We evaluate AI's role in CME diagnosis and its performance in "detection", "classification" and "segmentation" of OCT retinal images. We found that convolutional neural network (CNN)-based methods consistently outperformed other machine learning techniques, achieving an average accuracy of over 96% in detecting and identifying CME from OCT images. Despite certain limitations such as dataset size and ethical concerns, the synergy between AI and OCT, particularly through CNNs, holds promise for significantly advancing CME diagnostics.

PMID:38942125 | DOI:10.1016/j.survophthal.2024.06.005