Related Experiment Videos
EfficientNetB5-Based Deep Learning for Automated Cancer Detection in Tissue Microarray Images
Kuo-Wang Tsai1,2, Bach-Tung Pham3,4, Ching-Feng Cheng5,6,7
1Department of Research, Taipei Tzu Chi Hospital, Buddhist Tzu Chi Medical Foundation, New Taipei, Taiwan.
Background:
Tissue microarray (TMA) technology is pivotal in cancer research, enabling simultaneous analysis of multiple samples on a single slide. Manual analysis of TMA images remains time-consuming and subjective. This study aimed to develop a robust deep learning model to automate TMA image analysis, enhance accuracy, and support cancer diagnosis.
Methods:
We utilized the EfficientNetB5 deep learning architecture to classify TMA images from the DigitalSlide01 v1_1M01 dataset, comprising breast cancer, Upper Tract Urothelial Carcinoma (UTUC), and gastric cancer samples. Data augmentation methods, including Mix-Up and Cut-Mix, were employed to improve generalization. Preprocessing involved Otsu thresholding and smart bounding box cropping. The model was trained over 80 epochs using the Adam optimizer, a learning rate scheduler, and categorical cross-entropy loss. Performance was evaluated using 10-fold cross-validation and a separate test set, with accuracy, precision, recall, and F1-score as the primary metrics. Performance was further examined using physical-array-held-out and sample-level analyses.
Results:
Across the 10 cross-validation validation folds, mean accuracy and F1-score were 85.93% and 82.78%, respectively. Across the separate test-set evaluations by the ten fold-specific models, mean accuracy and F1-score were 80.08% and 75.13%, respectively. In the sample-level analysis, a 10-model mean-probability ensemble achieved 87.80% accuracy and 78.18% macro F1-score on the 246-core test subset. Cancer precision was 100.00% in the reported ensemble analysis, indicating that no Normal or Others samples were observed among samples predicted as Cancer; this does not imply that all Cancer samples were detected.
Conclusions:
The EfficientNetB5-based framework showed useful aggregate classification performance. These findings support the feasibility of automated TMA image classification and emphasize the need for source-balanced external multicenter validation before broader clinical application.