Bridging the gap between prostate radiology and pathology through machine learning

Abstract: Prostate cancer is the second deadliest cancer for American men. While Magnetic Resonance Imaging (MRI) is increasingly used to guide targeted biopsies for prostate cancer diagnosis, its utility remains limited due to high rates of false positives and false negatives as well as low inter-reader agreements. Machine learning methods to detect and localize cancer on prostate MRI can help standardize radiologist interpretations. However, existing machine learning methods vary not only in model architecture, but also in the ground truth labeling strategies used for model training. In this study, we compare different labeling strategies, namely, pathology-confirmed radiologist labels, pathologist labels on whole-mount histopathology images, and lesion-level and pixel-level digital pathologist labels (previously validated deep learning algorithm on histopathology images to predict pixel-level Gleason patterns) on whole-mount histopathology images. We analyse the effects these labels have on the performance of the trained machine learning models. Our experiments show that (1) radiologist labels and models trained with them can miss cancers, or underestimate cancer extent, (2) digital pathologist labels and models trained with them have high concordance with pathologist labels, and (3) models trained with digital pathologist labels achieve the best performance in prostate cancer detection in two different cohorts with different disease distributions, irrespective of the model architecture used. Digital pathologist labels can reduce challenges associated with human annotations, including labor, time, inter- and intra-reader variability, and can help bridge the gap between prostate radiology and pathology by enabling the training of reliable machine learning models to detect and localize prostate cancer on MRI.

• The paper demonstrates that using digital pathologist labels improves deep learning model accuracy and distinguishes between aggressive and indolent prostate cancers.
• It compares deep learning architectures such as SPCNet, U-Net, branched U-Net, and DeepLabv3+ trained on radiologist, pathologist, and digital pathology labels.
• The findings show that models trained with digital pathologist labels achieve higher Dice overlaps and ROC-AUC scores, indicating potential for more standardized diagnostics.

Bridging the Gap Between Prostate Radiology and Pathology Through Machine Learning

Introduction

Prostate cancer remains a significant clinical problem, ranking as the second deadliest cancer among American men. Magnetic Resonance Imaging (MRI) plays a crucial role in the detection and localization of prostate cancer, often guiding targeted biopsies. However, the high false positive and negative rates, coupled with challenges like inter-reader variability, limit the utility of MRI. This research investigates the optimization of prostate cancer detection on MRI using machine learning models, emphasizing the role of different training label types and introducing comparisons among them.

Methodology

Study Design

The study compared different labeling strategies to evaluate their impact on machine learning model performance for prostate cancer detection. Four deep learning models—SPCNet, U-Net, branched U-Net, and DeepLabv3+—were trained with datasets consisting of various label types, including pathology-confirmed radiologist labels, pathologist labels, and digital pathologist labels at both lesion and pixel levels.

Radiologist and pathologist labels were mapped onto MRI using automated registration platforms. The digital pathologist labels, derived from deep learning algorithms validated on histopathology images, offered a novel training path distinct from the globally utilized manual labeling.

Datasets and Labels

Data was acquired from 315 patients divided into radical prostatectomy and biopsy cohorts. The labeling included aggressive versus indolent cancer distinctions, essential for evaluating the machine learning models' precision in cancer identification.

Label Types

• Radiologist Labels ($\mathcal{L}^{Rad}$): MRI-based annotations confirmed by histopathological evidence.
• Pathologist Labels ($\mathcal{L}^{Path}$): Derived from histopathology images through registration onto MRI.
• Digital Pathologist Lesion Labels ($\mathcal{L}^{DPath}_{Lesion}$): Morphologically processed annotations differentiating cancer types.
• Digital Pathologist Pixel Labels ($\mathcal{L}^{DPath}_{Pixel}$): Provided precise pixel-level information, allowing a refined identification of cancerous components.

  Figure 1: Radiologists, pathologists or digital pathologists are used to create labels on MRI and serve to train deep learning models to detect cancer and aggressive cancer on MRI.

Results

Model Performance Evaluation

The SPCNet model consistently demonstrated superior performance across various label types. Digital pathologist labels enabled models to achieve higher or comparable accuracy compared to traditional radiologist or pathologist label-trained models. These models also uniquely facilitated the differentiation between aggressive and indolent cancer types, paving the way for more nuanced prostate cancer diagnostics.

Figure 2: Differences in labeling strategies in a typical patient in cohort C1-test (aggressive cancer - yellow, indolent cancer - green) showed on (a) T2w images and (b) ADC images...

Comparison of Labeling Strategies

Models trained on digital pathologist labels achieved superior Dice overlaps and lesion-level ROC-AUCs compared to radiologist-trained models. The pixel-level digital pathologist label models excelled at detecting mixed lesions within aggressive and indolent cancer contexts, showcasing an unprecedented discriminative ability not feasible with human annotations alone.

Figure 3: Predictions from SPCNet trained with different label types of a typical patient from cohort C1-test (same as Figure 2).

Discussion

The use of digital pathologist labels provides a robust methodological innovation. It not only challenges but potentially surpasses the conventional human expert label practices in prostate cancer detection via MRI. The findings suggest that these models can standardize and potentially enhance diagnostic accuracy while limiting the constraints of labor-intensive pathology labeling.

Furthermore, digital pathologist labels tackle several challenges, such as inter- and intra-reader variability and the intensive labor associated with human-generated annotations. They present a pathway to refine prostate cancer care by reliably distinguishing between aggressive and indolent forms, thus materially influencing treatment strategies.

Conclusion

Digital pathologist labels represent a significant advancement for training deep learning models in prostate cancer detection. This approach enhances model reliability and expands the potential for accurate, standardized prostate cancer diagnostics. Such innovations hold promise for improved clinical decision-making and patient outcomes in prostate cancer management.

Figure 4: Quantitative comparison between digital radiologist (SPCNet) predictions when trained and evaluated using different label types in cohort C1-test.