Genomic pan-cancer classification using image-based deep learning

AI support for colonoscopy quality control using CNN and transformer architectures

BACKGROUND

Construct deep learning models for colonoscopy quality control using different architectures and explore their decision-making mechanisms.

METHODS

A total of 4,189 colonoscopy images were collected from two medical centers, covering different levels of bowel cleanliness, the presence of polyps, and the cecum. Using these data, eight pre-trained models based on CNN and Transformer architectures underwent transfer learning and fine-tuning. The models' performance was evaluated using metrics such as AUC, Precision, and F1 score. Perceptual hash functions were employed to detect image changes, enabling real-time monitoring of colonoscopy withdrawal speed. Model interpretability was analyzed using techniques such as Grad-CAM and SHAP. Finally, the best-performing model was converted to ONNX format and deployed on device terminals.

RESULTS

The EfficientNetB2 model outperformed other architectures on the validation set, achieving an accuracy of 0.992. It surpassed models based on other CNN and Transformer architectures. The model's precision, recall, and F1 score were 0.991, 0.989, and 0.990, respectively. On the test set, the EfficientNetB2 model achieved an average AUC of 0.996, with a precision of 0.948 and a recall of 0.952. Interpretability analysis showed the specific image regions the model used for decision-making. The model was converted to ONNX format and deployed on device terminals, achieving an average inference speed of over 60 frames per second.

CONCLUSIONS

The AI-assisted quality system, based on the EfficientNetB2 model, integrates four key quality control indicators for colonoscopy. This integration enables medical institutions to comprehensively manage and enhance these indicators using a single model, showcasing promising potential for clinical applications.

Residual network improves the prediction accuracy of genomic selection

Genetic improvement of complex traits in animal and plant breeding depends on the efficient and accurate estimation of breeding values. Deep learning methods have been shown to be not superior over traditional genomic selection (GS) methods, partially due to the degradation problem (i.e. with the increase of the model depth, the performance of the deeper model deteriorates). Since the deep learning method residual network (ResNet) is designed to solve gradient degradation, we examined its performance and factors related to its prediction accuracy in GS. Here we compared the prediction accuracy of conventional genomic best linear unbiased prediction, Bayesian methods (BayesA, BayesB, BayesC, and Bayesian Lasso), and two deep learning methods, convolutional neural network and ResNet, on three datasets (wheat, simulated and real pig data). ResNet outperformed other methods in both Pearson's correlation coefficient (PCC) and mean squared error (MSE) on the wheat and simulated data. For the pig backfat depth trait, ResNet still had the lowest MSE, whereas Bayesian Lasso had the highest PCC. We further clustered the pig data into four groups and, on one separated group, ResNet had the highest prediction accuracy (both PCC and MSE). Transfer learning was adopted and capable of enhancing the performance of both convolutional neural network and ResNet. Taken together, our findings indicate that ResNet could improve GS prediction accuracy, affected potentially by factors such as the genetic architecture of complex traits, data volume, and heterogeneity.

PRADclass: Hybrid Gleason Grade-Informed Computational Strategy Identifies Consensus Biomarker Features Predictive of Aggressive Prostate Adenocarcinoma

BACKGROUND

Prostate adenocarcinoma (PRAD) is a common cancer diagnosis among men globally, yet large gaps in our knowledge persist with respect to the molecular bases of its progression and aggression. It is mostly indolent and slow-growing, but aggressive prostate cancers need to be recognized early for optimising treatment, with a view to reducing mortality.

METHODS

Based on TCGA transcriptomic data pertaining to PRAD and the associated clinical metadata, we determined the sample Gleason grade, and used it to execute: (i) Gleason-grade wise linear modeling, followed by five contrasts against controls and ten contrasts between grades; and (ii) Gleason-grade wise network modeling via weighted gene correlation network analysis (WGCNA). Candidate biomarkers were obtained from the above analysis and the consensus found. The consensus biomarkers were used as the feature space to train ML models for classifying a sample as benign, indolent or aggressive.

RESULTS

The statistical modeling yielded 77 Gleason grade-salient genes while the WGCNA algorithm yielded 1003 trait-specific key genes in grade-wise significant modules. Consensus analysis of the two approaches identified two genes in Grade-1 (SLC43A1 and PHGR1), 26 genes in Grade-4 (including LOC100128675, PPP1R3C, NECAB1, UBXN10, SERPINA5, CLU, RASL12, DGKG, FHL1, NCAM1, and CEND1), and seven genes in Grade-5 (CBX2, DPYS, FAM72B, SHCBP1, TMEM132A, TPX2, UBE2C). A RandomForest model trained and optimized on these 35 biomarkers for the ternary classification problem yielded a balanced accuracy ∼ 86% on external validation.

CONCLUSIONS

The consensus of multiple parallel computational strategies has unmasked candidate Gleason grade-specific biomarkers. PRADclass, a validated AI model featurizing these biomarkers achieved good performance, and could be trialed to predict the differentiation of prostate cancers. PRADclass is available for academic use at: https://apalania.shinyapps.io/pradclass (online) and https://github.com/apalania/pradclass (command-line interface).

Lung Cancer Subtype Diagnosis by Fusing Image-Genomics Data and Hybrid Deep Networks

Accurate diagnosis of cancer subtypes is crucial for precise treatment, because different cancer subtypes are involved with different pathology and require different therapies. Although deep learning techniques have made great success in computer vision and other fields, they do not work well on Lung cancer subtype diagnosis, due to the distinction of slide images between different cancer subtypes is ambiguous. Furthermore, they often over-fit to high-dimensional genomics data with limited samples, and do not fuse the image and genomics data in a sensible way. In this paper, we propose a hybrid deep network based approach LungDIG for Lung cancer subtype Diagnosis by fusing Image-Genomics data. LungDIG first tiles the tissue slide image into small patches and extracts the patch-level features by fine-tuning an Inception-V3 model. Since the patches may contain some false positives in non-diagnostic regions, it further designs a patch-level feature combination strategy to integrate the extracted patch features and maintain the diversity between different cancer subtypes. At the same time, it extracts the genomics features from Copy Number Variation data by an attention based nonlinear extractor. Next, it fuses the image and genomics features by an attention based multilayer perceptron (MLP) to diagnose cancer subtype. Experiments on TCGA lung cancer data show that LungDIG can not only achieve higher accuracy for cancer subtype diagnosis than state-of-the-art methods, but also have a high authenticity and good interpretability.

An integrated and automated testing approach on Inception Restnet-V3 based on convolutional neural network for leukocytes image classification

OBJECTIVES

The leukocyte is a specialized immune cell that functions as the foundation of the immune system and keeps the body healthy. The WBC classification plays a vital role in diagnosing various disorders in the medical area, including infectious diseases, immune deficiencies, leukemia, and COVID-19. A few decades ago, Machine Learning algorithms classified WBC types required for image segmentation, and the feature extraction stages, but this new approach becomes automatic while existing models can be fine-tuned for specific classifications.

METHODS

The inception architecture and deep learning model-based Resnet connection are integrated into this article. Our proposed method, inception Resnet-v3, was used to classify WBCs into five categories using 15.7k images. Pathologists made diagnoses of all images so a model could be trained to classify five distinct types of cells.

RESULTS

After implementing the proposed architecture on a large dataset of 5 categories of human peripheral white blood cells, it achieved high accuracy than VGG, U-Net and Resnet. We tested our model with WBC images from additional public datasets such as the Kaagel data sets and Raabin data sets of which the accuracy was 98.80% and 98.95%.

CONCLUSIONS

Considering the large sample sizes, we believe the proposed method can be used for improving the diagnostic performance of clinical blood examinations as well as a promising alternative for machine learning. Test results obtained with the system have been satisfying, with outstanding values for Accuracy, Precision, Recall, Specificity and F1 Score.