Challenges in Using Deep Neural Networks Across Multiple Readers in Delineating Prostate Gland Anatomy.

Area of Science:

• Medical image analysis
• Machine learning in oncology
• Radiology and medical imaging

Background:

• Deep learning models, such as deep neural networks (DNNs), show promise for automating complex medical image segmentation tasks.
• A significant challenge in training DNNs for oncological studies is the need for extensive data and diverse expert opinions, complicated by inter-reader variability.
• Inter-reader variability among clinical experts can hinder the reproducibility and generalization of DNN models in real-world clinical settings.

Purpose of the Study:

• To quantify the variability in DNN performance influenced by expert opinions in prostate gland segmentation.
• To explore strategies for training deep learning models and adapting them between different expert annotations.
• To improve the reproducibility of DNN models by addressing inter-reader variability in clinical expert annotations.

Main Methods:

• Utilized a 3D U-Net model for prostate gland segmentation on T2-weighted magnetic resonance imaging (MRI) data.
• Trained and tested the 3D U-Net model using data annotated by two expert readers (R#1 and R#2) individually and in combination.
• Evaluated model performance using the Dice coefficient and analyzed performance variations based on prostate gland size (large vs. small) via fivefold cross-validation.

Main Results:

• Individual training yielded average Dice coefficients of 0.825 for R#1 and 0.85 for R#2.
• Combined training with a representative proportion of data from both readers enhanced model reproducibility, achieving average Dice coefficients of 0.863 for R#1 and 0.869 for R#2.
• The model demonstrated improved performance on larger prostate glands (Dice: 0.846 for R#1, 0.872 for R#2) but showed diminished performance on smaller glands (Dice: 0.8 for R#1, 0.8 for R#2).

Conclusions:

• Combining expert annotations during the training of deep neural networks can significantly improve model reproducibility in medical image segmentation tasks.
• The proposed strategy effectively mitigates inter-reader variability, leading to more robust and generalizable DNN models for prostate cancer analysis.
• Future work should focus on strategies to improve DNN performance for segmenting smaller anatomical structures, which remain a challenge.