A deep dive into understanding tumor foci classification using multiparametric MRI based on convolutional neural...

Area of Science:

• Medical imaging analysis
• Machine learning in oncology
• Prostate cancer diagnostics

Background:

• Deep learning excels in disease classification with large datasets but faces challenges with limited medical imaging data, such as prostate multiparametric MRI (mpMRI).
• Interpreting deep learning models trained on medical images remains an under-explored area, hindering a full understanding of their predictive capabilities.
• Data scarcity and class imbalance are significant hurdles for applying deep learning to prostate mpMRI analysis.

Purpose of the Study:

• To develop an efficient convolutional neural network (CNN) for prostate cancer classification using mpMRI data.
• To systematically analyze the interpretation of CNNs at various convolutional layers to understand feature extraction and predictive power.
• To address the challenge of small sample sizes by integrating CNN-extracted features with a weighted extreme learning machine (wELM) classifier.

Main Methods:

• A CNN with four convolutional layers was trained on a retrospective dataset of 320 prostate MRIs from 201 patients (PROSTATEx Challenges 2017).
• The CNN was trained using tenfold cross-validation, incorporating T2-weighted (T2W), proton density-weighted (PD-W), dynamic contrast-enhanced (DCE), and diffusion-weighted (DW) imaging.
• Intermediate features from the CNN were extracted and fed into a wELM classifier, with model performance evaluated on a holdout cohort.

Main Results:

• The optimal input for the CNN was a combination of T2W, apparent diffusion coefficient (ADC), and DWI_b50 modalities.
• The CNN-CNN-wELM approach demonstrated superior performance compared to end-to-end CNN training.
• Feature map visualization showed lower CNN layers detecting basic features (edges, intensity changes) and higher layers learning abstract, lesion-specific features.

Conclusions:

• A novel workflow was established for analyzing small, imbalanced prostate mpMRI datasets by combining deep learning feature extraction with traditional machine learning classification.
• Model interpretation revealed that CNNs learn hierarchical features, progressing from low-level image characteristics to high-level diagnostic concepts.
• Saliency maps confirmed the model's ability to focus on relevant lesion regions while effectively filtering out background noise, aiding in prostate cancer patient stratification.