Automatic Grade Classification in Prostate Histopathological Images Using EfficientNet and Ordinal Focal Loss

Area of Science:

• Digital pathology
• Machine learning in oncology
• Computational imaging

Background:

• Accurate prostate cancer grading is crucial for treatment decisions.
• Existing automated methods struggle with high class similarity and label noise in ISUP grading.

Purpose of the Study:

• To develop an improved deep learning pipeline for automated International Society of Urological Pathology (ISUP) grade group classification.
• To address challenges in prostate histopathology image analysis, including class imbalance and label noise.

Main Methods:

• A pipeline using EfficientNet convolutional neural networks with a hybrid loss function (ordinal regression and Focal Loss).
• Implementation of a noise-filtering strategy based on prediction entropy to remove uncertain samples.
• Reformulation of the classification task as an ordinal regression problem to model grade hierarchy.

Main Results:

• Noise filtering improved performance from Cohen's kappa (κ) of 0.826 to 0.833.
• Incorporating ordinal loss further boosted performance to κ=0.851.
• The best model, combining ordinal regression and Focal Loss, achieved κ=0.857 and 0.669 accuracy, reducing severe misclassifications.

Conclusions:

• Explicitly modeling the ordinal structure of ISUP grades enhances classification accuracy.
• Mitigating label noise through prediction entropy is an effective strategy for improving automated prostate cancer grading.
• The proposed deep learning approach offers a promising tool for more accurate prostate cancer grading systems.