Multiple instance learning combined with label invariant synthetic data for guiding systematic prostate biopsy: a...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Prostate biopsy relies on ultrasound imaging with a systematic, non-targeted approach.
• This method is often blinded to intraprostatic pathology, leading to a high rate of false negatives.

Purpose of the Study:

• To develop a deep network for enhanced prostate cancer detection in systematic biopsies.
• To overcome challenges of statistical and limited labels in training data for improved diagnostic accuracy.

Main Methods:

• Utilized multiple instance learning (MIL) networks to learn from ultrasound image regions associated with statistical biopsy core pathology.
• Combined Independent Conditional Variational Auto Encoders (ICVAE) with MIL to generate synthetic data, alleviating the issue of limited biopsy samples.
• Trained ICVAE to learn label-invariant features from radiofrequency (RF) data for enhanced MIL network training.

Main Results:

• The study analyzed 339 prostate biopsy cores from 70 patients.
• Achieved an area under the curve (AUC) of 0.68.
• Reported sensitivity, specificity, and balanced accuracy of 0.77, 0.55, and 0.66, respectively.

Conclusions:

• The proposed deep learning approach offers a generic solution for improving cancer detection in systematic biopsies.
• This methodology is applicable to other scenarios with unlabeled data and noisy labels in training samples.