CNN Prostate Lesion Segmentation Using T2w and ADC MRI Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current diagnostic practices for prostate cancer, particularly using multi-parametric MRI, face challenges in accurately detecting and distinguishing between indolent and clinically significant prostate cancer due to inter-reader variability and suboptimal analysis, leading to under-detection and misclassification of lesions.
Innovation Solution
A system utilizing a convolutional neural network (CNN) that processes T2-weighted and apparent diffusion coefficient (ADC) data from multi-parametric MRI, employing focal loss and mutual finding loss to automatically segment and classify prostate lesions, predicting aggressiveness based on pixel-level likelihoods and providing indications of lesion presence and aggressiveness to users.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multi-parametric MRI with qualitative/semi-quantitative analysis (PI-RADS v2) is used for prostate cancer diagnosis, then anatomical and functional information is provided, but inter-reader variability and suboptimal analysis lead to under-detection and misclassification of lesions
Solution Approach 1:
The patent replaces the manual, expert-dependent qualitative/semi-quantitative analysis mechanism with an automated deep learning system. The CNN model processes mp-MRI data (T2w, ADC, DCE components) to generate pixel-level predictions, substituting human reader variability with consistent algorithmic evaluation. This automation resolves the contradiction by maintaining high detection accuracy while eliminating inter-reader variability through standardized processing of all input images.
Solution Approach 2:
The patent transforms the diagnostic approach from qualitative/semi-quantitative assessment to quantitative pixel-level probability outputs. By converting subjective radiologic findings into objective numerical predictions (likelihood values for each pixel), the system achieves both high precision and reliability. The loss function combines focal loss for hard samples and mutual information loss for multi-component integration, enabling quantitative diagnosis that overcomes the limitations of traditional qualitative methods.
2Measurement precision
If expert radiologic interpretation of mp-MRI is performed, then lesion detection capability is improved, but high level of expertise is required and analysis is time-consuming
Solution Approach 1:
The patent implements a self-service diagnostic system where the deep learning model autonomously processes mp-MRI data without requiring expert radiologist intervention for each case. The CNN model, trained on labeled datasets, performs automatic lesion detection and classification, generating diagnostic results that match or exceed expert-level performance. This eliminates the need for high-level expertise in routine interpretation while maintaining detection capability through the model's learned features from training data.
Solution Approach 2:
The patent performs preliminary analysis through automated processing of all mp-MRI components before clinical decision-making. The system pre-processes T2w, ADC, and DCE images through the CNN model to generate probability maps and lesion predictions, providing ready-to-use diagnostic information. This preliminary automated analysis reduces the time and expertise required for subsequent clinical evaluation, as the heavy lifting of lesion detection and characterization is already completed by the model.
3Ease of manufacture
If non-targeted template driven TRUS biopsy is used, then biopsy procedure is standardized, but under-detection of clinically significant PCa occurs
Solution Approach 1:
The patent performs preliminary lesion identification and localization using the deep learning model on mp-MRI data before biopsy execution. The system generates precise lesion boundaries and probability maps that guide targeted biopsy sampling, replacing non-targeted template-driven approaches. This preliminary anatomical and functional assessment enables clinicians to focus biopsy efforts on high-probability regions, significantly improving detection of clinically significant PCa while maintaining procedural standardization through consistent model-based guidance.
Data Source
AI summary
In accordance with some embodiments, systems, methods, and media for automatically segmenting and diagnosing prostate lesions using multi-parametric magnetic resonance imaging (mp-MRI) data are provided. In some embodiments, the system comprises is programmed to: receive mp-MRI data depicting a prostate, including T2w data and ADC data; provide the T2w data and ADC data as input to first and second input channels of a trained convolutional neural network (CNN); receive, from the trained CNN, output values from output channels indicating which pixels are likely to correspond to a particular class of prostate lesion, the channels corresponding to predicted aggressiveness in order of increasing aggressiveness, identify a prostate lesion in the data based on output values greater than a threshold; predict an aggressiveness based on which channel had values over the threshold; and present an indication that a prostate lesion of the predicted aggressiveness is likely present in the prostate.