Front Digit Health. 2024 Feb 28;6:1282043. doi: 10.3389/fdgth.2024.1282043. eCollection 2024.

ABSTRACT

Clinical narratives commonly use acronyms without explicitly defining their long forms. This makes it difficult to automatically interpret their sense as acronyms tend to be highly ambiguous. Supervised learning approaches to their disambiguation in the clinical domain are hindered by issues associated with patient privacy and manual annotation, which limit the size and diversity of training data. In this study, we demonstrate how scientific abstracts can be utilised to overcome these issues by creating a large automatically annotated dataset of artificially simulated global acronyms. A neural network trained on such a dataset achieved the F1-score of 95% on disambiguation of acronym mentions in scientific abstracts. This network was integrated with multi-word term recognition to extract a sense inventory of acronyms from a corpus of clinical narratives on the fly. Acronym sense extraction achieved the F1-score of 74% on a corpus of radiology reports. In clinical practice, the suggested approach can be used to facilitate development of institution-specific inventories.

PMID:38482049 | PMC:PMC10932973 | DOI:10.3389/fdgth.2024.1282043

Front Immunol. 2024 Feb 28;15:1352703. doi: 10.3389/fimmu.2024.1352703. eCollection 2024.

ABSTRACT

Deep learning models have been shown to accurately predict protein structure from sequence, allowing researchers to explore protein space from the structural viewpoint. In this paper we explore whether "novel" features, such as distinct loop conformations can arise from these predictions despite not being present in the training data. Here we have used ABodyBuilder2, a deep learning antibody structure predictor, to predict the structures of ~1.5M paired antibody sequences. We examined the predicted structures of the canonical CDR loops and found that most of these predictions fall into the already described CDR canonical form structural space. We also found a small number of "new" canonical clusters composed of heterogeneous sequences united by a common sequence motif and loop conformation. Analysis of these novel clusters showed their origins to be either shapes seen in the training data at very low frequency or shapes seen at high frequency but at a shorter sequence length. To evaluate explicitly the ability of ABodyBuilder2 to extrapolate, we retrained several models whilst withholding all antibody structures of a specific CDR loop length or canonical form. These "starved" models showed evidence of generalisation across CDRs of different lengths, but they did not extrapolate to loop conformations which were highly distinct from those present in the training data. However, the models were able to accurately predict a canonical form even if only a very small number of examples of that shape were in the training data. Our results suggest that deep learning protein structure prediction methods are unable to make completely out-of-domain predictions for CDR loops. However, in our analysis we also found that even minimal amounts of data of a structural shape allow the method to recover its original predictive abilities. We have made the ~1.5 M predicted structures used in this study available to download at https://doi.org/10.5281/zenodo.10280181.

PMID:38482007 | PMC:PMC10933040 | DOI:10.3389/fimmu.2024.1352703

Proc SPIE Int Soc Opt Eng. 2023 Jan-Feb;12391:1239106. doi: 10.1117/12.2650815. Epub 2023 Mar 16.

ABSTRACT

Whole slide imaging (WSI) is a common step used in histopathology to quickly digitize stained histological slides. Digital whole-slide images not only improve the efficiency of labeling but also open the door for computer-aided diagnosis, specifically machine learning-based methods. Hyperspectral imaging (HSI) is an imaging modality that captures data in various wavelengths, some beyond the range of visible lights. In this study, we developed and implemented an automated microscopy system that can acquire hyperspectral whole slide images (HWSI). The system is robust since it consists of parts that can be swapped and bought from different manufacturers. We used the automated system and built a database of 49 HWSI of thyroid cancer. The automatic whole-slide hyperspectral imaging microscope can have many potential applications in biological and medical areas.

PMID:38481979 | PMC:PMC10932749 | DOI:10.1117/12.2650815

Front Cardiovasc Med. 2024 Feb 28;11:1354517. doi: 10.3389/fcvm.2024.1354517. eCollection 2024.

ABSTRACT

OBJECTIVE: This study aims to review the application of deep learning techniques in the imaging diagnosis and treatment of aortic aneurysm (AA), focusing on screening, diagnosis, lesion segmentation, surgical assistance, and prognosis prediction.

METHODS: A comprehensive literature review was conducted, analyzing studies that utilized deep learning models such as Convolutional Neural Networks (CNNs) in various aspects of AA management. The review covered applications in screening, segmentation, surgical planning, and prognosis prediction, with a focus on how these models improve diagnosis and treatment outcomes.

RESULTS: Deep learning models demonstrated significant advancements in AA management. For screening and diagnosis, models like ResNet achieved high accuracy in identifying AA in non-contrast CT scans. In segmentation, techniques like U-Net provided precise measurements of aneurysm size and volume, crucial for surgical planning. Deep learning also assisted in surgical procedures by accurately predicting stent placement and postoperative complications. Furthermore, models were able to predict AA progression and patient prognosis with high accuracy.

CONCLUSION: Deep learning technologies show remarkable potential in enhancing the diagnosis, treatment, and management of AA. These advancements could lead to more accurate and personalized patient care, improving outcomes in AA management.

PMID:38481955 | PMC:PMC10933089 | DOI:10.3389/fcvm.2024.1354517

Front Artif Intell. 2024 Feb 28;7:1326050. doi: 10.3389/frai.2024.1326050. eCollection 2024.

ABSTRACT

Covert tobacco advertisements often raise regulatory measures. This paper presents that artificial intelligence, particularly deep learning, has great potential for detecting hidden advertising and allows unbiased, reproducible, and fair quantification of tobacco-related media content. We propose an integrated text and image processing model based on deep learning, generative methods, and human reinforcement, which can detect smoking cases in both textual and visual formats, even with little available training data. Our model can achieve 74% accuracy for images and 98% for text. Furthermore, our system integrates the possibility of expert intervention in the form of human reinforcement. Using the pre-trained multimodal, image, and text processing models available through deep learning makes it possible to detect smoking in different media even with few training data.

PMID:38481821 | PMC:PMC10936563 | DOI:10.3389/frai.2024.1326050

Ann Gastroenterol. 2024 Mar-Apr;37(2):133-141. doi: 10.20524/aog.2024.0861. Epub 2024 Feb 14.

ABSTRACT

Integrating artificial intelligence (AI) into gastrointestinal (GI) endoscopy heralds a significant leap forward in managing GI disorders. AI-enabled applications, such as computer-aided detection and computer-aided diagnosis, have significantly advanced GI endoscopy, improving early detection, diagnosis and personalized treatment planning. AI algorithms have shown promise in the analysis of endoscopic data, critical in conditions with traditionally low diagnostic sensitivity, such as indeterminate biliary strictures and pancreatic cancer. Convolutional neural networks can markedly improve the diagnostic process when integrated with cholangioscopy or endoscopic ultrasound, especially in the detection of malignant biliary strictures and cholangiocarcinoma. AI's capacity to analyze complex image data and offer real-time feedback can streamline endoscopic procedures, reduce the need for invasive biopsies, and decrease associated adverse events. However, the clinical implementation of AI faces challenges, including data quality issues and the risk of overfitting, underscoring the need for further research and validation. As the technology matures, AI is poised to become an indispensable tool in the gastroenterologist's arsenal, necessitating the integration of robust, validated AI applications into routine clinical practice. Despite remarkable advances, challenges such as operator-dependent accuracy and the need for intricate examinations persist. This review delves into the transformative role of AI in enhancing endoscopic diagnostic accuracy, particularly highlighting its utility in the early detection and personalized treatment of GI diseases.

PMID:38481787 | PMC:PMC10927620 | DOI:10.20524/aog.2024.0861

Oral Surg Oral Med Oral Pathol Oral Radiol. 2024 Feb 12:S2212-4403(24)00062-2. doi: 10.1016/j.oooo.2024.02.003. Online ahead of print.

ABSTRACT

OBJECTIVE: We examined the effectiveness and feasibility of the Mask Region-based Convolutional Neural Network (Mask R-CNN) for automatic detection of cephalometric landmarks on lateral cephalometric radiographs (LCRs).

STUDY DESIGN: In total, 400 LCRs, each with 19 manually identified landmarks, were collected. Of this total, 320 images were randomly selected as the training dataset for Mask R-CNN, and the remaining 80 images were used for testing the automatic detection of the 19 cephalometric landmarks, for a total of 1520 landmarks. Detection rate, average error, and detection accuracy rate were calculated to assess Mask R-CNN performance.

RESULTS: Of the 1520 landmarks, 1494 were detected, for a detection rate of 98.29%. The average error, or linear deviation distance between the detected points and the originally marked points of each detected landmark, ranged from 0.56 to 9.51 mm, with an average of 2.19 mm. For detection accuracy rate, 649 landmarks (43.44%) had a linear deviation distance less than 1 mm, 1020 (68.27%) less than 2 mm, and 1281 (85.74%) less than 4 mm in deviation from the manually marked point. The average detection time was 1.48 seconds per image.

CONCLUSIONS: Deep learning Mask R-CNN shows promise in enhancing cephalometric analysis by automating landmark detection on LCRs, addressing the limitations of manual analysis, and demonstrating effectiveness and feasibility.

PMID:38480069 | DOI:10.1016/j.oooo.2024.02.003

J Cardiovasc Comput Tomogr. 2024 Mar 12:S1934-5925(24)00060-1. doi: 10.1016/j.jcct.2024.03.001. Online ahead of print.

ABSTRACT

BACKGROUND: ECG-gated cardiac CT is now widely used in infants with congenital heart disease (CHD). Deep Learning Image Reconstruction (DLIR) could improve image quality while minimizing the radiation dose.

OBJECTIVES: To define the potential dose reduction using DLIR with an anthropomorphic phantom.

METHOD: An anthropomorphic pediatric phantom was scanned with an ECG-gated cardiac CT at four dose levels. Images were reconstructed with an iterative and a deep-learning reconstruction algorithm (ASIR-V and DLIR). Detectability of high-contrast vessels were computed using a mathematical observer. Discrimination between two vessels was assessed by measuring the CT spatial resolution. The potential dose reduction while keeping a similar level of image quality was assessed.

RESULTS: DLIR-H enhances detectability by 2.4% and discrimination performances by 20.9% in comparison with ASIR-V 50. To maintain a similar level of detection, the dose could be reduced by 64% using high-strength DLIR in comparison with ASIR-V50.

CONCLUSION: DLIR offers the potential for a substantial dose reduction while preserving image quality compared to ASIR-V.

PMID:38480035 | DOI:10.1016/j.jcct.2024.03.001

Curr Probl Cancer. 2024 Mar 13:101077. doi: 10.1016/j.currproblcancer.2024.101077. Online ahead of print.

ABSTRACT

Skin cancer, including the highly lethal malignant melanoma, poses a significant global health challenge with a rising incidence rate. Early detection plays a pivotal role in improving survival rates. This study aims to develop an advanced deep learning-based approach for accurate skin lesion classification, addressing challenges such as limited data availability, class imbalance, and noise. Modern deep neural network designs, such as ResNeXt101, SeResNeXt101, ResNet152V2, DenseNet201, GoogLeNet, and Xception, which are used in the study and ze optimised using the SGD technique. The dataset comprises diverse skin lesion images from the HAM10000 and ISIC datasets. Noise and artifacts are tackled using image inpainting, and data augmentation techniques enhance training sample diversity. The ensemble technique is utilized, creating both average and weighted average ensemble models. Grid search optimizes model weight distribution. The individual models exhibit varying performance, with metrics including recall, precision, F1 score, and MCC. The "Average ensemble model" achieves harmonious balance, emphasizing precision, F1 score, and recall, yielding high performance. The "Weighted ensemble model" capitalizes on individual models' strengths, showcasing heightened precision and MCC, yielding outstanding performance. The ensemble models consistently outperform individual models, with the average ensemble model attaining a macro-average ROC-AUC score of 96 % and the weighted ensemble model achieving a macro-average ROC-AUC score of 97 %. This research demonstrates the efficacy of ensemble techniques in significantly improving skin lesion classification accuracy. By harnessing the strengths of individual models and addressing their limitations, the ensemble models exhibit robust and reliable performance across various metrics. The findings underscore the potential of ensemble techniques in enhancing medical diagnostics and contributing to improved patient outcomes in skin lesion diagnosis.

PMID:38480028 | DOI:10.1016/j.currproblcancer.2024.101077

Trends Biotechnol. 2024 Mar 13:S0167-7799(24)00038-6. doi: 10.1016/j.tibtech.2024.02.009. Online ahead of print.

ABSTRACT

In pathology and biomedical research, histology is the cornerstone method for tissue analysis. Currently, the histological workflow consumes plenty of chemicals, water, and time for staining procedures. Deep learning is now enabling digital replacement of parts of the histological staining procedure. In virtual staining, histological stains are created by training neural networks to produce stained images from an unstained tissue image, or through transferring information from one stain to another. These technical innovations provide more sustainable, rapid, and cost-effective alternatives to traditional histological pipelines, but their development is in an early phase and requires rigorous validation. In this review we cover the basic concepts of virtual staining for histology and provide future insights into the utilization of artificial intelligence (AI)-enabled virtual histology.

PMID:38480025 | DOI:10.1016/j.tibtech.2024.02.009

J UOEH. 2024;46(1):119-122. doi: 10.7888/juoeh.46.119.

ABSTRACT

There is clear scientific evidence that physical activity helps to prevent depression and anxiety. Utilizing mobile health (mHealth) technologies to enable physical activity is promising, but the evidence of the effectiveness of mHealth interventions on physical activity and mental health is inconsistent. We recently developed a native smartphone app to prevent depression and anxiety by promoting physical activity. One of the app's strengths is that it adopts a deep-learning model and automatically estimates psychological distress from users' physical activity patterns. We conducted a single-arm, 1-month feasibility trial to examine the implementation of the app and its effectiveness in promoting physical activity and improving depression and anxiety. As a result, we did not observe any significant improvement in physical activity or psychological distress. For implementation aspects, the participants used the app less. The conclusion of the presentation is that mHealth interventions are promising for the improvement of physical activity and mental health among workers, but, at this stage, their effectiveness is unclear. There are challenges to be addressed, especially in implementation.

PMID:38479866 | DOI:10.7888/juoeh.46.119

Int J Cardiol. 2024 Mar 11:131945. doi: 10.1016/j.ijcard.2024.131945. Online ahead of print.

ABSTRACT

BACKGROUND: Quantitative coronary angiography (QCA) offers objective and reproducible measures of coronary lesions. However, significant inter- and intra-observer variability and time-consuming processes hinder the practical application of on-site QCA in the current clinical setting. This study proposes a novel method for artificial intelligence-based QCA (AI-QCA) analysis of the major vessels and evaluates its performance.

METHODS: AI-QCA was developed using three deep-learning models trained on 7658 angiographic images from 3129 patients for the precise delineation of lumen boundaries. An automated quantification method, employing refined matching for accurate diameter calculation and iterative updates of diameter trend lines, was embedded in the AI-QCA. A separate dataset of 676 coronary angiography images from 370 patients was retrospectively analyzed to compare AI-QCA with manual QCA performed by expert analysts. A match was considered between manual and AI-QCA lesions when the minimum lumen diameter (MLD) location identified manually coincided with the location identified by AI-QCA. Matched lesions were evaluated in terms of diameter stenosis (DS), MLD, reference lumen diameter (RLD), and lesion length (LL).

RESULTS: AI-QCA exhibited a sensitivity of 89% in lesion detection and strong correlations with manual QCA for DS, MLD, RLD, and LL. Among 995 matched lesions, most cases (892 cases, 80%) exhibited DS differences ≤10%. Multiple lesions of the major vessels were accurately identified and quantitatively analyzed without manual corrections.

CONCLUSION: AI-QCA demonstrates promise as an automated tool for analysis in coronary angiography, offering potential advantages for the quantitative assessment of coronary lesions and clinical decision-making.

PMID:38479496 | DOI:10.1016/j.ijcard.2024.131945

Neurobiol Dis. 2024 Mar 11:106472. doi: 10.1016/j.nbd.2024.106472. Online ahead of print.

ABSTRACT

BACKGROUND: Whether there is hypothalamic degeneration in Parkinson's disease (PD) and its association with clinical symptoms and pathophysiological changes remains controversial.

OBJECTIVES: We aimed to quantify microstructural changes in hypothalamus using a novel deep learning-based tool in patients with PD and those with probable rapid-eye-movement sleep behavior disorder (pRBD). We further assessed whether these microstructural changes associated with clinical symptoms and free thyroxine (FT4) levels.

METHODS: This study included 186 PD, 67 pRBD, and 179 healthy controls. Multi-shell diffusion MRI were scanned and mean kurtosis (MK) in hypothalamic subunits were calculated. Participants were assessed using Unified Parkinson's Disease Rating Scale (UPDRS), RBD Questionnaire-Hong Kong (RBDQ-HK), Hamilton Depression Rating Scale (HAMD), and Activity of Daily Living (ADL) Scale. Additionally, a subgroup of PD (n = 31) underwent assessment of FT4.

RESULTS: PD showed significant decreases of MK in anterior-superior (a-sHyp), anterior-inferior (a-iHyp), superior tubular (supTub), and inferior tubular hypothalamus when compared with healthy controls. Similarly, pRBD exhibited decreases of MK in a-iHyp and supTub. In PD group, MK in above four subunits were significantly correlated with UPDRS-I, HAMD, and ADL. Moreover, MK in a-iHyp and a-sHyp were significantly correlated with FT4 level. In pRBD group, correlations were observed between MK in a-iHyp and UPDRS-I.

CONCLUSIONS: Our study reveals that microstructural changes in the hypothalamus are already significant at the early neurodegenerative stage. These changes are associated with emotional alterations, daily activity levels, and thyroid hormone levels.

PMID:38479482 | DOI:10.1016/j.nbd.2024.106472

Radiother Oncol. 2024 Mar 11:110221. doi: 10.1016/j.radonc.2024.110221. Online ahead of print.

ABSTRACT

BACKGROUND AND PURPOSE: To develop a computed tomography (CT)-based deep learning model to predict overall survival (OS) among small-cell lung cancer (SCLC) patients and identify patients who could benefit from prophylactic cranial irradiation (PCI) based on OS signature risk stratification.

MATERIALS AND METHODS: This study retrospectively included 556 SCLC patients from three medical centers. The training, internal validation, and external validation cohorts comprised 309, 133, and 114 patients, respectively. The OS signature was built using a unified fully connected neural network. A deep learning model was developed based on the OS signature. Clinical and combined models were developed and compared with a deep learning model. Additionally, the benefits of PCI were evaluated after stratification using an OS signature.

RESULTS: Within the internal and external validation cohorts, the deep learning model (concordance index [C-index] 0.745, 0.733) was far superior to the clinical model (C-index: 0.635, 0.630) in predicting OS, but slightly worse than the combined model (C-index: 0.771, 0.770). Additionally, the deep learning model had excellent calibration, clinical usefulness, and improved accuracy in classifying survival outcomes. Remarkably, patients at high risk had a survival benefit from PCI in both the limited and extensive stages (all P < 0.05), whereas no significant association was observed in patients at low risk.

CONCLUSIONS: The CT-based deep learning model exhibited promising performance in predicting the OS of SCLC patients. The OS signature may aid in individualized treatment planning to select patients who may benefit from PCI.

PMID:38479441 | DOI:10.1016/j.radonc.2024.110221

Comput Methods Programs Biomed. 2024 Mar 8;248:108113. doi: 10.1016/j.cmpb.2024.108113. Online ahead of print.

ABSTRACT

BACKGROUND AND OBJECTIVE: In recent years, Artificial Intelligence (AI) and in particular Deep Neural Networks (DNN) became a relevant research topic in biomedical image segmentation due to the availability of more and more data sets along with the establishment of well known competitions. Despite the popularity of DNN based segmentation on the research side, these techniques are almost unused in the daily clinical practice even if they could support effectively the physician during the diagnostic process. Apart from the issues related to the explainability of the predictions of a neural model, such systems are not integrated in the diagnostic workflow, and a standardization of their use is needed to achieve this goal.

METHODS: This paper presents IODeep a new DICOM Information Object Definition (IOD) aimed at storing both the weights and the architecture of a DNN already trained on a particular image dataset that is labeled as regards the acquisition modality, the anatomical region, and the disease under investigation.

RESULTS: The IOD architecture is presented along with a DNN selection algorithm from the PACS server based on the labels outlined above, and a simple PACS viewer purposely designed for demonstrating the effectiveness of the DICOM integration, while no modifications are required on the PACS server side. Also a service based architecture in support of the entire workflow has been implemented.

CONCLUSION: IODeep ensures full integration of a trained AI model in a DICOM infrastructure, and it is also enables a scenario where a trained model can be either fine-tuned with hospital data or trained in a federated learning scheme shared by different hospitals. In this way AI models can be tailored to the real data produced by a Radiology ward thus improving the physician decision making process. Source code is freely available at https://github.com/CHILab1/IODeep.git.

PMID:38479148 | DOI:10.1016/j.cmpb.2024.108113

Comput Methods Programs Biomed. 2024 Mar 7;248:108111. doi: 10.1016/j.cmpb.2024.108111. Online ahead of print.

ABSTRACT

BACKGROUND AND OBJECTIVE: Training deep learning models for medical image segmentation require large annotated datasets, which can be expensive and time-consuming to create. Active learning is a promising approach to reduce this burden by strategically selecting the most informative samples for segmentation. This study investigates the use of active learning for efficient left ventricle segmentation in echocardiography with sparse expert annotations.

METHODS: We adapt and evaluate various sampling techniques, demonstrating their effectiveness in judiciously selecting samples for segmentation. Additionally, we introduce a novel strategy, Optimised Representativeness Sampling, which combines feature-based outliers with the most representative samples to enhance annotation efficiency.

RESULTS: Our findings demonstrate a substantial reduction in annotation costs, achieving a remarkable 99% upper bound performance while utilising only 20% of the labelled data. This equates to a reduction of 1680 images needing annotation within our dataset. When applied to a publicly available dataset, our approach yielded a remarkable 70% reduction in required annotation efforts, representing a significant advancement compared to baseline active learning strategies, which achieved only a 50% reduction. Our experiments highlight the nuanced performance of diverse sampling strategies across datasets within the same domain.

CONCLUSIONS: The study provides a cost-effective approach to tackle the challenges of limited expert annotations in echocardiography. By introducing a distinct dataset, made publicly available for research purposes, our work contributes to the field's understanding of efficient annotation strategies in medical image segmentation.

PMID:38479147 | DOI:10.1016/j.cmpb.2024.108111

Biomol Concepts. 2024 Mar 13;15(1). doi: 10.1515/bmc-2022-0043. eCollection 2024 Jan 1.

ABSTRACT

Computational biology faces many challenges like protein secondary structure prediction (PSS), prediction of solvent accessibility, etc. In this work, we addressed PSS prediction. PSS is based on sequence-structure mapping and interaction among amino acid residues. We proposed an encoder-decoder with an attention mechanism model, which considers the mapping of sequence structure and interaction among residues. The attention mechanism is used to select prominent features from amino acid residues. The proposed model is trained on CB513 and CullPDB open datasets using the Nvidia DGX system. We have tested our proposed method for Q 3 and Q 8 accuracy, segment of overlap, and Mathew correlation coefficient. We achieved 70.63 and 78.93% Q 3 and Q 8 accuracy, respectively, on the CullPDB dataset whereas 79.8 and 77.13% Q 3 and Q 8 accuracy on the CB513 dataset. We observed improvement in SOV up to 80.29 and 91.3% on CullPDB and CB513 datasets. We achieved the results using our proposed model in very few epochs, which is better than the state-of-the-art methods.

PMID:38478635 | DOI:10.1515/bmc-2022-0043

PLoS Comput Biol. 2024 Mar 13;20(3):e1011074. doi: 10.1371/journal.pcbi.1011074. Online ahead of print.

ABSTRACT

Deep learning is a powerful tool for neural decoding, broadly applied to systems neuroscience and clinical studies. Interpretable and transparent models that can explain neural decoding for intended behaviors are crucial to identifying essential features of deep learning decoders in brain activity. In this study, we examine the performance of deep learning to classify mouse behavioral states from mesoscopic cortex-wide calcium imaging data. Our convolutional neural network (CNN)-based end-to-end decoder combined with recurrent neural network (RNN) classifies the behavioral states with high accuracy and robustness to individual differences on temporal scales of sub-seconds. Using the CNN-RNN decoder, we identify that the forelimb and hindlimb areas in the somatosensory cortex significantly contribute to behavioral classification. Our findings imply that the end-to-end approach has the potential to be an interpretable deep learning method with unbiased visualization of critical brain regions.

PMID:38478563 | DOI:10.1371/journal.pcbi.1011074

PLoS One. 2024 Mar 13;19(3):e0297536. doi: 10.1371/journal.pone.0297536. eCollection 2024.

ABSTRACT

Nasal endoscopy is routinely performed to distinguish the pathological types of masses. There is a lack of studies on deep learning algorithms for discriminating a wide range of endoscopic nasal cavity mass lesions. Therefore, we aimed to develop an endoscopic-examination-based deep learning model to detect and classify nasal cavity mass lesions, including nasal polyps (NPs), benign tumors, and malignant tumors. The clinical feasibility of the model was evaluated by comparing the results to those of manual assessment. Biopsy-confirmed nasal endoscopic images were obtained from 17 hospitals in South Korea. Here, 400 images were used for the test set. The training and validation datasets consisted of 149,043 normal nasal cavity, 311,043 NP, 9,271 benign tumor, and 5,323 malignant tumor lesion images. The proposed Xception architecture achieved an overall accuracy of 0.792 with the following class accuracies on the test set: normal = 0.978 ± 0.016, NP = 0.790 ± 0.016, benign = 0.708 ± 0.100, and malignant = 0.698 ± 0.116. With an average area under the receiver operating characteristic curve (AUC) of 0.947, the AUC values and F1 score were highest in the order of normal, NP, malignant tumor, and benign tumor classes. The classification performances of the proposed model were comparable with those of manual assessment in the normal and NP classes. The proposed model outperformed manual assessment in the benign and malignant tumor classes (sensitivities of 0.708 ± 0.100 vs. 0.549 ± 0.172, 0.698 ± 0.116 vs. 0.518 ± 0.153, respectively). In urgent (malignant) versus nonurgent binary predictions, the deep learning model achieved superior diagnostic accuracy. The developed model based on endoscopic images achieved satisfactory performance in classifying four classes of nasal cavity mass lesions, namely normal, NP, benign tumor, and malignant tumor. The developed model can therefore be used to screen nasal cavity lesions accurately and rapidly.

PMID:38478548 | DOI:10.1371/journal.pone.0297536

PLoS One. 2024 Mar 13;19(3):e0299970. doi: 10.1371/journal.pone.0299970. eCollection 2024.

ABSTRACT

The accuracy of traditional CT image segmentation algorithms is hindered by issues such as low contrast and high noise in the images. While numerous scholars have introduced deep learning-based CT image segmentation algorithms, they still face challenges, particularly in achieving high edge accuracy and addressing pixel classification errors. To tackle these issues, this study proposes the MIS-Net (Medical Images Segment Net) model, a deep learning-based approach. The MIS-Net model incorporates multi-scale atrous convolution into the encoding and decoding structure with symmetry, enabling the comprehensive extraction of multi-scale features from CT images. This enhancement aims to improve the accuracy of lung and liver edge segmentation. In the evaluation using the COVID-19 CT Lung and Infection Segmentation dataset, the left and right lung segmentation results demonstrate that MIS-Net achieves a Dice Similarity Coefficient (DSC) of 97.61. Similarly, in the Liver Tumor Segmentation Challenge 2017 public dataset, the DSC of MIS-Net reaches 98.78.

PMID:38478519 | DOI:10.1371/journal.pone.0299970