Front Plant Sci. 2024 Mar 14;15:1367828. doi: 10.3389/fpls.2024.1367828. eCollection 2024.

ABSTRACT

Precise and timely leaf area index (LAI) estimation for winter wheat is crucial for precision agriculture. The emergence of high-resolution unmanned aerial vehicle (UAV) data and machine learning techniques offers a revolutionary approach for fine-scale estimation of wheat LAI at the low cost. While machine learning has proven valuable for LAI estimation, there are still model limitations and variations that impede accurate and efficient LAI inversion. This study explores the potential of classical machine learning models and deep learning model for estimating winter wheat LAI using multispectral images acquired by drones. Initially, the texture features and vegetation indices served as inputs for the partial least squares regression (PLSR) model and random forest (RF) model. Then, the ground-measured LAI data were combined to invert winter wheat LAI. In contrast, this study also employed a convolutional neural network (CNN) model that solely utilizes the cropped original image for LAI estimation. The results show that vegetation indices outperform the texture features in terms of correlation analysis with LAI and estimation accuracy. However, the highest accuracy is achieved by combining both vegetation indices and texture features to invert LAI in both conventional machine learning methods. Among the three models, the CNN approach yielded the highest LAI estimation accuracy (R 2 = 0.83), followed by the RF model (R 2 = 0.82), with the PLSR model exhibited the lowest accuracy (R 2 = 0.78). The spatial distribution and values of the estimated results for the RF and CNN models are similar, whereas the PLSR model differs significantly from the first two models. This study achieves rapid and accurate winter wheat LAI estimation using classical machine learning and deep learning methods. The findings can serve as a reference for real-time wheat growth monitoring and field management practices.

PMID:38550285 | PMC:PMC10972960 | DOI:10.3389/fpls.2024.1367828

Front Nucl Med. 2023;3:1326549. doi: 10.3389/fnume.2023.1326549. Epub 2023 Nov 10.

NO ABSTRACT

PMID:38550275 | PMC:PMC10976900 | DOI:10.3389/fnume.2023.1326549

Front Neuroimaging. 2024 Mar 14;3:1349415. doi: 10.3389/fnimg.2024.1349415. eCollection 2024.

ABSTRACT

Diffusion magnetic resonance imaging is sensitive to the microstructural properties of brain tissue. However, estimating clinically and scientifically relevant microstructural properties from the measured signals remains a highly challenging inverse problem that machine learning may help solve. This study investigated if recently developed rotationally invariant spherical convolutional neural networks can improve microstructural parameter estimation. We trained a spherical convolutional neural network to predict the ground-truth parameter values from efficiently simulated noisy data and applied the trained network to imaging data acquired in a clinical setting to generate microstructural parameter maps. Our network performed better than the spherical mean technique and multi-layer perceptron, achieving higher prediction accuracy than the spherical mean technique with less rotational variance than the multi-layer perceptron. Although we focused on a constrained two-compartment model of neuronal tissue, the network and training pipeline are generalizable and can be used to estimate the parameters of any Gaussian compartment model. To highlight this, we also trained the network to predict the parameters of a three-compartment model that enables the estimation of apparent neural soma density using tensor-valued diffusion encoding.

PMID:38550242 | PMC:PMC10972853 | DOI:10.3389/fnimg.2024.1349415

J Nucl Med. 2024 Mar 28:jnumed.123.266761. doi: 10.2967/jnumed.123.266761. Online ahead of print.

ABSTRACT

Heart failure (HF) is a leading cause of morbidity and mortality in the United States and worldwide, with a high associated economic burden. This study aimed to assess whether artificial intelligence models incorporating clinical, stress test, and imaging parameters could predict hospitalization for acute HF exacerbation in patients undergoing SPECT/CT myocardial perfusion imaging. Methods: The HF risk prediction model was developed using data from 4,766 patients who underwent SPECT/CT at a single center (internal cohort). The algorithm used clinical risk factors, stress variables, SPECT imaging parameters, and fully automated deep learning-generated calcium scores from attenuation CT scans. The model was trained and validated using repeated hold-out (10-fold cross-validation). External validation was conducted on a separate cohort of 2,912 patients. During a median follow-up of 1.9 y, 297 patients (6%) in the internal cohort were admitted for HF exacerbation. Results: The final model demonstrated a higher area under the receiver-operating-characteristic curve (0.87 ± 0.03) for predicting HF admissions than did stress left ventricular ejection fraction (0.73 ± 0.05, P < 0.0001) or a model developed using only clinical parameters (0.81 ± 0.04, P < 0.0001). These findings were confirmed in the external validation cohort (area under the receiver-operating-characteristic curve: 0.80 ± 0.04 for final model, 0.70 ± 0.06 for stress left ventricular ejection fraction, 0.72 ± 0.05 for clinical model; P < 0.001 for all). Conclusion: Integrating SPECT myocardial perfusion imaging into an artificial intelligence-based risk assessment algorithm improves the prediction of HF hospitalization. The proposed method could enable early interventions to prevent HF hospitalizations, leading to improved patient care and better outcomes.

PMID:38548351 | DOI:10.2967/jnumed.123.266761

J Arthroplasty. 2024 Mar 26:S0883-5403(24)00268-7. doi: 10.1016/j.arth.2024.03.053. Online ahead of print.

ABSTRACT

BACKGROUND: Dissatisfaction after total knee arthroplasty (TKA) ranges from 15 to 30%. While patient selection may be partially responsible, morphological and reconstructive challenges may be determinants. Preoperative computed tomography (CT) scans for TKA planning allow us to evaluate the hip-knee-ankle axis and establish a baseline phenotypic distribution across anatomic parameters. The purpose of this cross-sectional analysis was to establish the distributions of 27 parameters in a pre-TKA cohort and perform threshold analysis to identify anatomic outliers.

METHODS: There were 1,352 pre-TKA CTs that were processed. A two-step deep learning pipeline of classification and segmentation models identified landmark images, then generated contour representations. We utilized an open-source computer vision library to compute measurements for 27 anatomic metrics along the hip-knee axis. Normative distribution plots were established, and thresholds for the 15th percentile at both extremes were calculated. Metrics falling outside the central 70th percentile were considered outlier indices. A threshold analysis of outlier indices against the proportion of the cohort was performed.

RESULTS: Significant variation exists in pre-TKA anatomy across 27 normally distributed metrics. Threshold analysis revealed a sigmoid function with a critical point at nine outlier indices, representing 31.2% of subjects as anatomic outliers. Metrics with the greatest variation related to deformity (tibiofemoral angle, medial proximal tibial angle, lateral distal femoral angle), bony size (tibial width, anteroposterior femoral size, femoral head size, medial femoral condyle size), intraoperative landmarks (posterior tibial slope, transepicondylar and posterior condylar axes), and neglected rotational considerations (acetabular and femoral version, femoral torsion).

CONCLUSION: In the largest non-industry database of pre-TKA CTs using a fully automated three-stage deep learning and computer vision-based pipeline, marked anatomic variation exists. In the pursuit of understanding the dissatisfaction rate after TKA, acknowledging that 31% of patients represent anatomic outliers may help us better achieve anatomically personalized TKA, with or without adjunctive technology.

PMID:38548237 | DOI:10.1016/j.arth.2024.03.053

J Dent. 2024 Mar 26:104971. doi: 10.1016/j.jdent.2024.104971. Online ahead of print.

ABSTRACT

OBJECTIVES: In many prosthodontic procedures, traditional computer-aided design (CAD) is often time-consuming and lacks accuracy in shape restoration. In this study, we innovatively combined implicit template and deep learning (DL) to construct a precise neural network for personalized tooth defect restoration.

METHODS: Ninety models of right maxillary central incisor (80 for training, 10 for validation) were collected. A DL model named ToothDIT was trained to establish an implicit template and a neural network capable of predicting unique identifications. In the validation stage, teeth in validation set were processed into corner, incisive, and medium defects. The defective teeth were inputted into ToothDIT to predict the unique identification, which actuated the deformation of the implicit template to generate the highly customized template (DIT) for the target tooth. Morphological restorations were executed with templates from template shape library (TSL), average tooth template (ATT), and DIT in Exocad (GmbH, Germany). RMSestimate, width, length, aspect ratio, incisal edge curvature, incisive end retraction, and guiding inclination were introduced to assess the restorative accuracy. Statistical analysis was conducted using two-way ANOVA and paired t-test for overall and detailed differences.

RESULTS: DIT displayed significantly smaller RMSestimate than TSL and ATT. In 2D detailed analysis, DIT exhibited significantly less deviations from the natural teeth compared to TSL and ATT.

CONCLUSION: The proposed DL model successfully reconstructed the morphology of anterior teeth with various degrees of defects and achieved satisfactory accuracy. This approach provides a more reliable reference for prostheses design, resulting in enhanced accuracy in morphological restoration.

CLINICAL SIGNIFICANCE: This DL model holds promise in assisting dentists and technicians in obtaining morphology templates that closely resemble the original shape of the defective teeth. These customized templates serve as a foundation for enhancing the efficiency and precision of digital restorative design for defective teeth.

PMID:38548165 | DOI:10.1016/j.jdent.2024.104971

J Stomatol Oral Maxillofac Surg. 2024 Mar 26:101840. doi: 10.1016/j.jormas.2024.101840. Online ahead of print.

ABSTRACT

OBJECTIVE: To conduct a systematic review with meta-analyses to assess the recent scientific literature addressing the application of deep learning radiomics in oral squamous cell carcinoma (OSCC).

MATERIALS AND METHODS: Electronic and manual literature retrieval was performed using PubMed, Web of Science, EMbase, Ovid-MEDLINE, and IEEE databases from 2012 to 2023. The ROBINS-I tool was used for quality evaluation; random-effects model was used; and results were reported according to the PRISMA statement.

RESULTS: A total of 26 studies involving 64,731 medical images were included in quantitative synthesis. The meta-analysis showed that, the pooled sensitivity and specificity were 0.88 (95%CI: 0.87∼0.88) and 0.80 (95%CI: 0.80∼0.81), respectively. Deeks' asymmetry test revealed there existed slight publication bias (P = 0.03).

CONCLUSIONS: The advances in the application of radiomics combined with learning algorithm in OSCC were reviewed, including diagnosis and differential diagnosis of OSCC, efficacy assessment and prognosis prediction. The demerits of deep learning radiomics at the current stage and its future development direction aimed at medical imaging diagnosis were also summarized and analyzed at the end of the article.

PMID:38548062 | DOI:10.1016/j.jormas.2024.101840

Micron. 2024 Mar 17;181:103633. doi: 10.1016/j.micron.2024.103633. Online ahead of print.

ABSTRACT

This study utilized X-ray micro-computed tomography (micro-CT) to investigate weathered gypsum rocks which can or do serve as a rock substrate for endolithic organisms, focusing on their internal pore-fracture microstructure, estimating porosity, and quantitative comparison between various samples. Examining sections and reconstructed 3D models provides a more detailed insight into the overall structural conditions within rock fragments and the interconnectivity in pore networks, surpassing the limitations of analyzing individual 2D images. Results revealed diverse gypsum forms, cavities, fractures, and secondary features influenced by weathering. Using deep learning segmentation based on the U-Net models within the Dragonfly software enabled to identify and visualize the porous systems and determinate void space which was used to calculate porosity. This approach allowed to describe what type of microstructures and cavities is responsible for the porous spaces in different gypsum samples. A set of quantitative analysis of the detected void and modeled networks provided a needed information about the development of the pore system, connectivity, and pore size distribution. Comparison with mercury intrusion porosimetry showed that both methods consider different populations of pores. In our case, micro-CT typically detects larger pores (> 10 μm) which is related to the effective resolution of the scanned images. Still, micro-CT demonstrated to be an efficient tool in examining the internal microstructures of weathered gypsum rocks, with promising implications particularly in geobiology and microbiology for the characterization of lithic habitats.

PMID:38547790 | DOI:10.1016/j.micron.2024.103633

Comput Biol Med. 2024 Mar 26;173:108385. doi: 10.1016/j.compbiomed.2024.108385. Online ahead of print.

ABSTRACT

Alkaliphilic proteins have great potential as biocatalysts in biotechnology, especially for enzyme engineering. Extensive research has focused on exploring the enzymatic potential of alkaliphiles and characterizing alkaliphilic proteins. However, the current method employed for identifying these proteins that requires web lab experiment is time-consuming, labor-intensive, and expensive. Therefore, the development of a computational method for alkaliphilic protein identification would be invaluable for protein engineering and design. In this study, we present a novel approach that uses embeddings from a protein language model called ESM-2(3B) in a deep learning framework to classify alkaliphilic and non-alkaliphilic proteins. To our knowledge, this is the first attempt to employ embeddings from a pre-trained protein language model to classify alkaliphilic protein. A reliable dataset comprising 1,002 alkaliphilic and 1,866 non-alkaliphilic proteins was constructed for training and testing the proposed model. The proposed model, dubbed ALPACA, achieves performance scores of 0.88, 0.84, and 0.75 for accuracy, f1-score, and Matthew correlation coefficient respectively on independent dataset. ALPACA is likely to serve as a valuable resource for exploring protein alkalinity and its role in protein design and engineering.

PMID:38547659 | DOI:10.1016/j.compbiomed.2024.108385

Comput Biol Med. 2024 Mar 18;173:108339. doi: 10.1016/j.compbiomed.2024.108339. Online ahead of print.

ABSTRACT

The application of Artificial Intelligence (AI) to screen drug molecules with potential therapeutic effects has revolutionized the drug discovery process, with significantly lower economic cost and time consumption than the traditional drug discovery pipeline. With the great power of AI, it is possible to rapidly search the vast chemical space for potential drug-target interactions (DTIs) between candidate drug molecules and disease protein targets. However, only a small proportion of molecules have labelled DTIs, consequently limiting the performance of AI-based drug screening. To solve this problem, a machine learning-based approach with great ability to generalize DTI prediction across molecules is desirable. Many existing machine learning approaches for DTI identification failed to exploit the full information with respect to the topological structures of candidate molecules. To develop a better approach for DTI prediction, we propose GraphormerDTI, which employs the powerful Graph Transformer neural network to model molecular structures. GraphormerDTI embeds molecular graphs into vector-format representations through iterative Transformer-based message passing, which encodes molecules' structural characteristics by node centrality encoding, node spatial encoding and edge encoding. With a strong structural inductive bias, the proposed GraphormerDTI approach can effectively infer informative representations for out-of-sample molecules and as such, it is capable of predicting DTIs across molecules with an exceptional performance. GraphormerDTI integrates the Graph Transformer neural network with a 1-dimensional Convolutional Neural Network (1D-CNN) to extract the drugs' and target proteins' representations and leverages an attention mechanism to model the interactions between them. To examine GraphormerDTI's performance for DTI prediction, we conduct experiments on three benchmark datasets, where GraphormerDTI achieves a superior performance than five state-of-the-art baselines for out-of-molecule DTI prediction, including GNN-CPI, GNN-PT, DeepEmbedding-DTI, MolTrans and HyperAttentionDTI, and is on a par with the best baseline for transductive DTI prediction. The source codes and datasets are publicly accessible at https://github.com/mengmeng34/GraphormerDTI.

PMID:38547658 | DOI:10.1016/j.compbiomed.2024.108339

Comput Biol Med. 2024 Mar 20;173:108280. doi: 10.1016/j.compbiomed.2024.108280. Online ahead of print.

ABSTRACT

BACKGROUND: Timely detection of neurodevelopmental and neurological conditions is crucial for early intervention. Specific Language Impairment (SLI) in children and Parkinson's disease (PD) manifests in speech disturbances that may be exploited for diagnostic screening using recorded speech signals. We were motivated to develop an accurate yet computationally lightweight model for speech-based detection of SLI and PD, employing novel feature engineering techniques to mimic the adaptable dynamic weight assignment network capability of deep learning architectures.

MATERIALS AND METHODS: In this research, we have introduced an advanced feature engineering model incorporating a novel feature extraction function, the Factor Lattice Pattern (FLP), which is a quantum-inspired method and uses a superposition-like mechanism, making it dynamic in nature. The FLP encompasses eight distinct patterns, from which the most appropriate pattern was discerned based on the data structure. Through the implementation of the FLP, we automatically extracted signal-specific textural features. Additionally, we developed a new feature engineering model to assess the efficacy of the FLP. This model is self-organizing, producing nine potential results and subsequently choosing the optimal one. Our speech classification framework consists of (1) feature extraction using the proposed FLP and a statistical feature extractor; (2) feature selection employing iterative neighborhood component analysis and an intersection-based feature selector; (3) classification via support vector machine and k-nearest neighbors; and (4) outcome determination using combinational majority voting to select the most favorable results.

RESULTS: To validate the classification capabilities of our proposed feature engineering model, designed to automatically detect PD and SLI, we employed three speech datasets of PD and SLI patients. Our presented FLP-centric model achieved classification accuracy of more than 95% and 99.79% for all PD and SLI datasets, respectively.

CONCLUSIONS: Our results indicate that the proposed model is an accurate alternative to deep learning models in classifying neurological conditions using speech signals.

PMID:38547655 | DOI:10.1016/j.compbiomed.2024.108280

Biomed Tech (Berl). 2024 Mar 29. doi: 10.1515/bmt-2024-0028. Online ahead of print.

ABSTRACT

Leukemia is a class of hematologic malignancies, of which acute myeloid leukemia (AML) is the most common. Screening and diagnosis of AML are performed by microscopic examination or chemical testing of images of the patient's peripheral blood smear. In smear-microscopy, the ability to quickly identify, count, and differentiate different types of blood cells is critical for disease diagnosis. With the development of deep learning (DL), classification techniques based on neural networks have been applied to the recognition of blood cells. However, DL methods have high requirements for the number of valid datasets. This study aims to assess the applicability of the auxiliary classification generative adversarial network (ACGAN) in the classification task for small samples of white blood cells. The method is trained on the TCIA dataset, and the classification accuracy is compared with two classical classifiers and the current state-of-the-art methods. The results are evaluated using accuracy, precision, recall, and F1 score. The accuracy of the ACGAN on the validation set is 97.1 % and the precision, recall, and F1 scores on the validation set are 97.5 , 97.3, and 97.4 %, respectively. In addition, ACGAN received a higher score in comparison with other advanced methods, which can indicate that it is competitive in classification accuracy.

PMID:38547466 | DOI:10.1515/bmt-2024-0028

PLoS Comput Biol. 2024 Mar 28;20(3):e1012004. doi: 10.1371/journal.pcbi.1012004. Online ahead of print.

ABSTRACT

In evolutionary models, mutations are exogenously introduced by the modeler, rather than endogenously introduced by the replicator itself. We present a new deep-learning based computational model, the self-replicating artificial neural network (SeRANN). We train it to (i) copy its own genotype, like a biological organism, which introduces endogenous spontaneous mutations; and (ii) simultaneously perform a classification task that determines its fertility. Evolving 1,000 SeRANNs for 6,000 generations, we observed various evolutionary phenomena such as adaptation, clonal interference, epistasis, and evolution of both the mutation rate and the distribution of fitness effects of new mutations. Our results demonstrate that universal evolutionary phenomena can naturally emerge in a self-replicator model when both selection and mutation are implicit and endogenous. We therefore suggest that SeRANN can be applied to explore and test various evolutionary dynamics and hypotheses.

PMID:38547320 | DOI:10.1371/journal.pcbi.1012004

PLoS One. 2024 Mar 28;19(3):e0300444. doi: 10.1371/journal.pone.0300444. eCollection 2024.

ABSTRACT

This paper presents a novel sound event detection (SED) system for rare events occurring in an open environment. Wavelet multiresolution analysis (MRA) is used to decompose the input audio clip of 30 seconds into five levels. Wavelet denoising is then applied on the third and fifth levels of MRA to filter out the background. Significant transitions, which may represent the onset of a rare event, are then estimated in these two levels by combining the peak-finding algorithm with the K-medoids clustering algorithm. The small portions of one-second duration, called 'chunks' are cropped from the input audio signal corresponding to the estimated locations of the significant transitions. Features from these chunks are extracted by the wavelet scattering network (WSN) and are given as input to a support vector machine (SVM) classifier, which classifies them. The proposed SED framework produces an error rate comparable to the SED systems based on convolutional neural network (CNN) architecture. Also, the proposed algorithm is computationally efficient and lightweight as compared to deep learning models, as it has no learnable parameter. It requires only a single epoch of training, which is 5, 10, 200, and 600 times lesser than the models based on CNNs and deep neural networks (DNNs), CNN with long short-term memory (LSTM) network, convolutional recurrent neural network (CRNN), and CNN respectively. The proposed model neither requires concatenation with previous frames for anomaly detection nor any additional training data creation needed for other comparative deep learning models. It needs to check almost 360 times fewer chunks for the presence of rare events than the other baseline systems used for comparison in this paper. All these characteristics make the proposed system suitable for real-time applications on resource-limited devices.

PMID:38547253 | DOI:10.1371/journal.pone.0300444

PLoS One. 2024 Mar 28;19(3):e0300725. doi: 10.1371/journal.pone.0300725. eCollection 2024.

ABSTRACT

Named Entity Recognition (NER) is a natural language processing task that has been widely explored for different languages in the recent decade but is still an under-researched area for the Urdu language due to its rich morphology and language complexities. Existing state-of-the-art studies on Urdu NER use various deep-learning approaches through automatic feature selection using word embeddings. This paper presents a deep learning approach for Urdu NER that harnesses FastText and Floret word embeddings to capture the contextual information of words by considering the surrounding context of words for improved feature extraction. The pre-trained FastText and Floret word embeddings are publicly available for Urdu language which are utilized to generate feature vectors of four benchmark Urdu language datasets. These features are then used as input to train various combinations of Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), Gated Recurrent Unit (GRU), CRF, and deep learning models. The results show that our proposed approach significantly outperforms existing state-of-the-art studies on Urdu NER, achieving an F-score of up to 0.98 when using BiLSTM+GRU with Floret embeddings. Error analysis shows a low classification error rate ranging from 1.24% to 3.63% across various datasets showing the robustness of the proposed approach. The performance comparison shows that the proposed approach significantly outperforms similar existing studies.

PMID:38547173 | DOI:10.1371/journal.pone.0300725

Proc Natl Acad Sci U S A. 2024 Apr 2;121(14):e2320013121. doi: 10.1073/pnas.2320013121. Epub 2024 Mar 28.

ABSTRACT

Dephosphorylation of pSer51 of the α subunit of translation initiation factor 2 (eIF2αP) terminates signaling in the integrated stress response (ISR). A trimeric mammalian holophosphatase comprised of a protein phosphatase 1 (PP1) catalytic subunit, the conserved C-terminally located ~70 amino acid core of a substrate-specific regulatory subunit (PPP1R15A/GADD34 or PPP1R15B/CReP) and G-actin (an essential cofactor) efficiently dephosphorylate eIF2αP in vitro. Unlike their viral or invertebrate counterparts, with whom they share the conserved 70 residue core, the mammalian PPP1R15s are large proteins of more than 600 residues. Genetic and cellular observations point to a functional role for regions outside the conserved core of mammalian PPP1R15A in dephosphorylating its natural substrate, the eIF2 trimer. We have combined deep learning technology, all-atom molecular dynamics simulations, X-ray crystallography, and biochemistry to uncover binding of the γ subunit of eIF2 to a short helical peptide repeated four times in the functionally important N terminus of human PPP1R15A that extends past its conserved core. Binding entails insertion of Phe and Trp residues that project from one face of an α-helix formed by the conserved repeats of PPP1R15A into a hydrophobic groove exposed on the surface of eIF2γ in the eIF2 trimer. Replacing these conserved Phe and Trp residues with Ala compromises PPP1R15A function in cells and in vitro. These findings suggest mechanisms by which contacts between a distant subunit of eIF2 and elements of PPP1R15A distant to the holophosphatase active site contribute to dephosphorylation of eIF2αP by the core PPP1R15 holophosphatase and to efficient termination of the ISR in mammals.

PMID:38547060 | DOI:10.1073/pnas.2320013121

IEEE Trans Med Imaging. 2024 Mar 28;PP. doi: 10.1109/TMI.2024.3382909. Online ahead of print.

ABSTRACT

Non-invasive prostate cancer classification from MRI has the potential to revolutionize patient care by providing early detection of clinically significant disease, but has thus far shown limited positive predictive value. To address this, we present a image-based deep learning method to predict clinically significant prostate cancer from screening MRI in patients that subsequently underwent biopsy with results ranging from benign pathology to the highest grade tumors. Specifically, we demonstrate that mixed supervision via diverse histopathological ground truth improves classification performance despite the cost of reduced concordance with image-based segmentation. Where prior approaches have utilized pathology results as ground truth derived from targeted biopsies and whole-mount prostatectomy to strongly supervise the localization of clinically significant cancer, our approach also utilizes weak supervision signals extracted from nontargeted systematic biopsies with regional localization to improve overall performance. Our key innovation is performing regression by distribution rather than simply by value, enabling use of additional pathology findings traditionally ignored by deep learning strategies. We evaluated our model on a dataset of 973 (testing n = 198) multi-parametric prostate MRI exams collected at UCSF from 2016-2019 followed by MRI/ultrasound fusion (targeted) biopsy and systematic (nontargeted) biopsy of the prostate gland, demonstrating that deep networks trained with mixed supervision of histopathology can feasibly exceed the performance of the Prostate Imaging-Reporting and Data System (PI-RADS) clinical standard for prostate MRI interpretation (71.6% vs 66.7% balanced accuracy and 0.724 vs 0.716 AUC).

PMID:38547000 | DOI:10.1109/TMI.2024.3382909

IEEE Trans Image Process. 2024 Mar 28;PP. doi: 10.1109/TIP.2024.3378461. Online ahead of print.

ABSTRACT

Existing deep learning methods for fine-grained visual recognition often rely on large-scale, well-annotated training data. Obtaining fine-grained annotations in the wild typically requires concentration and expertise, such as fine category annotation for species recognition, instance annotation for person re-identification (re-id) and dense annotation for segmentation, which inevitably leads to label noise. This paper aims to tackle label noise in deep model training for fine-grained visual recognition. We propose a Neighbor-Attention Label Correction (NALC) model to correct labels during the training stage. NALC samples a training batch and a validation batch from the training set. It hence leverages a meta-learning framework to correct labels in the training batch based on the validation batch. To enhance the optimization efficiency, we introduce a novel nested optimization algorithm for the meta-learning framework. The proposed training procedure consistently improves label accuracy in the training batch, consequently enhancing the learned image representation. Experimental results demonstrate that our method significantly increases label accuracy from 70% to over 98% and outperforms recent approaches by up to 13.4% in mean Average Precision (mAP) on various fine-grained image retrieval (FGIR) tasks, including instance retrieval on CUB200 and person re-id on Market1501. We also demonstrate the efficacy of NALC on noisy semantic segmentation datasets generated from Cityscapes, where it achieves a significant 7.8% improvement in mIOU score. NALC also exhibits robustness to different types of noise, including simulated noise such as Asymmetric, Pair-Flip, and Pattern noise, as well as practical noisy labels generated by tracklets and clustering.

PMID:38546993 | DOI:10.1109/TIP.2024.3378461

IEEE Trans Neural Netw Learn Syst. 2024 Mar 28;PP. doi: 10.1109/TNNLS.2024.3359275. Online ahead of print.

ABSTRACT

Variational autoencoders (VAEs) are challenged by the imbalance between representation inference and task fitting caused by surrogate loss. To address this issue, existing methods adjust their balance by directly tuning their coefficients. However, these methods suffer from a tradeoff uncertainty, i.e., nondynamic regulation over iterations and inflexible hyperparameters for learning tasks. Accordingly, we make the first attempt to introduce an evolutionary VAE (eVAE), building on the variational information bottleneck (VIB) theory and integrative evolutionary neural learning. eVAE integrates a variational genetic algorithm (VGA) into VAE with variational evolutionary operators, including variational mutation (V-mutation), crossover, and evolution. Its training mechanism synergistically and dynamically addresses and updates the learning tradeoff uncertainty in the evidence lower bound (ELBO) without additional constraints and hyperparameter tuning. Furthermore, eVAE presents an evolutionary paradigm to tune critical factors of VAEs and addresses the premature convergence and random search problem in integrating evolutionary optimization into deep learning. Experiments show that eVAE addresses the KL-vanishing problem for text generation with low reconstruction loss, generates all the disentangled factors with sharp images, and improves image generation quality. eVAE achieves better disentanglement, generation performance, and generation-inference balance than its competitors. Code available at: https://github.com/amasawa/eVAE.

PMID:38546992 | DOI:10.1109/TNNLS.2024.3359275

J Appl Clin Med Phys. 2024 Mar 28:e14350. doi: 10.1002/acm2.14350. Online ahead of print.

ABSTRACT

OBJECTIVE: Adaptive planning to accommodate anatomic changes during treatment often requires repeated segmentation. In this study, prior patient-specific data was integrateda into a registration-guided multi-channel multi-path (Rg-MCMP) segmentation framework to improve the accuracy of repeated clinical target volume (CTV) segmentation.

METHODS: This study was based on CT image datasets for a total of 90 cervical cancer patients who received two courses of radiotherapy. A total of 15 patients were selected randomly as the test set. In the Rg-MCMP segmentation framework, the first-course CT images (CT1) were registered to second-course CT images (CT2) to yield aligned CT images (aCT1), and the CTV in the first course (CTV1) was propagated to yield aligned CTV contours (aCTV1). Then, aCT1, aCTV1, and CT2 were combined as the inputs for 3D U-Net consisting of a channel-based multi-path feature extraction network. The performance of the Rg-MCMP segmentation framework was evaluated and compared with the single-channel single-path model (SCSP), the standalone registration methods, and the registration-guided multi-channel single-path (Rg-MCSP) model. The Dice similarity coefficient (DSC), 95% Hausdorff distance (HD95), and average surface distance (ASD) were used as the metrics.

RESULTS: The average DSC of CTV for the deformable image DIR-MCMP model was found to be 0.892, greater than that of the standalone DIR (0.856), SCSP (0.837), and DIR-MCSP (0.877), which were improvements of 4.2%, 6.6%, and 1.7%, respectively. Similarly, the rigid body DIR-MCMP model yielded an average DSC of 0.875, which exceeded standalone RB (0.787), SCSP (0.837), and registration-guided multi-channel single-path (0.848), which were improvements of 11.2%, 4.5%, and 3.2%, respectively. These improvements in DSC were statistically significant (p < 0.05).

CONCLUSION: The proposed Rg-MCMP framework achieved excellent accuracy in CTV segmentation as part of the adaptive radiotherapy workflow.

PMID:38546277 | DOI:10.1002/acm2.14350