Assessing the Performance of Deep Learning for Automated Gleason Grading in Prostate Cancer

Area of Science:

• Digital pathology
• Artificial intelligence
• Oncology

Background:

• Prostate cancer diagnosis relies heavily on accurate Gleason grading.
• Current grading methods can be subjective and time-consuming.
• Advanced computational tools are needed for efficient and precise prostate cancer assessment.

Purpose of the Study:

• To compare the efficacy of 11 deep neural network (DNN) architectures for automated Gleason grading in prostate carcinoma.
• To evaluate the performance of traditional versus recent DNN architectures in classifying prostate tissue samples.
• To identify the most promising AI models for improving prostate cancer diagnostics.

Main Methods:

• Utilized a dataset of 34,264 annotated prostate tissue tiles.
• Employed a standardized image classification pipeline based on the AUCMEDI framework.
• Evaluated 11 distinct deep neural network architectures, including both traditional and recent models.

Main Results:

• Significant variation in sensitivity was observed across different DNN architectures.
• Recent architectures generally outperformed traditional ones in automated Gleason grading.
• The ConvNeXt model demonstrated the highest performance, balancing complexity and generalizability.
• Challenges remain in differentiating between closely related Gleason grades.

Conclusions:

• Deep neural networks, especially newer architectures like ConvNeXt, show significant potential for automated Gleason grading.
• AI-driven digital pathology can enhance the accuracy and efficiency of prostate cancer diagnosis.
• Further research into differentiating subtle Gleason grade variations is warranted.