Semisupervised Learning with Report-guided Pseudo Labels for Deep Learning-based Prostate Cancer Detection Using...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Deep learning models for malignancy detection require large annotated datasets, which are costly and time-consuming to acquire.
• Prostate cancer diagnosis relies heavily on MRI, but accurate interpretation can be challenging.
• Semisupervised learning (SSL) offers a way to leverage unlabeled data, but its performance can be limited.

Purpose of the Study:

• To evaluate a novel report-guided semisupervised learning (RG-SSL) method for detecting clinically significant prostate cancer.
• To assess the efficiency of RG-SSL in leveraging automated sparse information from diagnostic reports for deep learning.
• To compare the performance of RG-SSL against traditional supervised learning (SL) and other SSL methods.

Main Methods:

• Developed a report-guided semisupervised learning (RG-SSL) method using biparametric MRI for prostate cancer detection.
• Trained RG-SSL, SL, and state-of-the-art SSL methods on varying numbers of manually annotated MRI examinations (100-3050).
• Compared model performance on an external dataset using receiver operating characteristic (ROC) and free-response ROC analysis.

Main Results:

• At 100 annotations, RG-SSL achieved a higher diagnostic AUC (0.86) than SL (0.78) and best SSL (0.81).
• RG-SSL matched the examination-based performance of SL using 14 times fewer annotations (169 vs. 3050).
• Lesion-based performance of RG-SSL with 431 annotations was equivalent to SL with 3050 annotations.

Conclusions:

• RG-SSL significantly outperforms standard SSL in detecting clinically significant prostate cancer.
• RG-SSL achieves performance comparable to supervised learning with substantially reduced annotation requirements.
• This method offers a highly efficient approach for deep learning-based prostate cancer detection, improving annotation efficiency.