Prostate cancer (PCa) requires precise tumor localization on imaging for radiation therapy planning, active surveillance, and focal treatment targeting. Multiparametric MRI (mpMRI) is the clinical standard for prostate imaging, but manual delineation of tumor boundaries is time-consuming and highly variable between and within readers, motivating the development of automated CNN-based segmentation tools.
A fundamental challenge for AI-based prostate tumor segmentation is the absence of verified ground truth. Biopsy confirms cancer only at the specific needle sites sampled, not across the entire prostate. MRI-drawn tumor contours by radiologists or oncologists are imperfect reference standards, as they may systematically underestimate true tumor volumes relative to actual cancer distribution. Without a reliable ground truth, it is difficult to know whether an AI segmentation is correct or simply mirroring the biases in its training labels.
A second barrier to clinical adoption is that CNN predictions are typically opaque: the network produces a segmentation output, but clinicians cannot directly see which image features drove the prediction. This lack of interpretability reduces trust and makes it difficult to identify when a model is reasoning from clinically meaningful tissue patterns versus confounding artifacts or noise.
This study addressed both barriers simultaneously: it validated a prostate tumor segmentation CNN against whole-mount histopathology (the most accurate ground truth available, obtained from surgically removed prostates) and applied an explainability method (Grad-CAM) to generate heat maps revealing which image regions the network used to make its segmentation decisions.
The study used mpMRI data from 137 patients with histologically confirmed prostate cancer, acquired between 2008 and 2019 on 1.5 Tesla and 3 Tesla Siemens MRI systems at the University of Freiburg Medical Center. MRI sequences included T2-weighted images, diffusion-weighted imaging (DWI) for ADC map generation, and dynamic contrast-enhanced (DCE) imaging. A synthetic high b-value DWI image (b = 1400 s/mm2) was calculated for each patient to standardize diffusion weighting across scanner field strengths.
The 137 patients were divided into a training cohort of 122 patients (including radiation therapy and prostatectomy patients) and a test cohort of 15 prostatectomy patients for whom whole-mount histopathology was available. For the test cohort, the surgically removed prostates were sliced, stained with hematoxylin and eosin, and expert-annotated to delineate cancer regions. These histology slices were then registered to the corresponding pre-operative MRI images using multi-step anatomical landmark-based co-registration, creating matched in-vivo MRI and histopathology tumor labels for the same patient.
This co-registration process provided two independent types of ground truth for the test cohort: PCa-Rad (tumor contours drawn by two experienced radiation oncologists directly on the MRI using PI-RADS criteria) and PCa-Histo (tumor contours from the histopathology, registered to the MRI). This allowed the study to directly compare CNN performance against both the clinical standard and the pathological gold standard in the same patients.
The CNN used a 3D U-Net architecture with three encoder and three decoder blocks, trained on 64x64x16 voxel patches extracted from the MRI volumes. Class-balanced sampling (equal probability of background, prostate gland, and cancer patches) was used to address the severe imbalance between the small tumor volumes and large amounts of normal tissue in each scan. Training was conducted for 150 epochs with Bayesian hyperparameter optimization.
Gradient-weighted Class Activation Mapping (Grad-CAM) is an explainability technique that generates a spatial heat map showing which parts of the input image were most important for a neural network's prediction. It works by computing the gradients of the class score with respect to the feature maps in the last convolutional layer of the CNN, then weighting those feature maps proportionally and projecting the result back onto the input image dimensions.
The authors extended the standard Grad-CAM method -- originally developed for image classification -- to the segmentation task. In segmentation, the prediction is not a single class score but a per-voxel classification. The adaptation averages the class scores across all voxels within the predicted segmentation mask before computing gradients, effectively asking: which input regions were most influential in producing the overall segmentation of that tissue class?
The resulting heat maps highlight the spatial regions within the MRI that the CNN relied on most when predicting each structure. For a trustworthy prostate cancer segmentation model, the Grad-CAM map should concentrate activation within or near the tumor region -- not in anatomically unrelated areas. This provides clinicians with a visual sanity check to assess whether the network is reasoning from genuine tissue contrast patterns.
A model randomization test was conducted as a mathematical sanity check: heat maps generated from the trained CNN were compared with heat maps from a completely untrained network with random weights. A meaningful explainability method should produce very different outputs between trained and untrained networks. The metric used was mutual information (MI) and structural similarity index (SSIM) between the two sets of heat maps.
The CNN achieved a mean Dice Similarity Coefficient (DSC) of 0.62 for prostate gland segmentation -- consistent with published literature -- and mean DSC values of 0.31 against the histopathology ground truth and 0.32 against the radiologist-drawn ground truth for prostate cancer segmentation. These tumor DSC values are low in absolute terms but are consistent with the 0.35 reported in comparable studies and closely match the published human interobserver variability range of 0.48 to 0.52 for prostate cancer delineation on MRI.
Critically, the DSC values comparing CNN predictions with the histopathology ground truth and with the radiologist ground truth were not significantly different from each other. This means the CNN performed no worse when evaluated against the true pathological cancer distribution than it did against the clinical MRI contours -- suggesting the network learned genuine tissue discrimination rather than simply mimicking radiologist contour styles.
In five cases, the CNN produced very poor segmentation (DSC = 0 to 0.1). Investigation of these cases revealed one had residual bleeding from a prior biopsy altering tissue appearance, and four showed pronounced benign prostatic hyperplasia (BPH) -- enlarged inner gland tissue that can visually mimic cancer on MRI. These represent known confounders for both human and AI prostate cancer detection on imaging.
The Grad-CAM heat maps showed the highest intersection over union (IOU) with the ground truth at a threshold of 0.5, confirming that the network attention was genuinely concentrated at cancer locations. The model randomization test confirmed the heat maps reflect learned patterns: MI and SSIM values between the trained and untrained networks were 1 and 0 respectively -- meaning the trained and random models produced completely different heat maps, validating that the trained model's explanations encode real learned information.
The relatively low absolute DSC for tumor segmentation reflects a fundamental problem: PI-RADS-based MRI contours -- used as training labels -- systematically underestimate true tumor volumes compared with histopathology. A CNN trained on these labels inherits their underestimation bias. This creates a ceiling on achievable performance that is unrelated to the neural network's capacity and cannot be overcome without better training annotations.
The near-identical CNN performance against both ground truth types (radiologist and histology) suggests the network learned features consistent with actual cancer biology rather than idiosyncrasies of individual readers. That the CNN converged on predictions that disagree comparably with both reference standards indicates the algorithm may have found an intermediate solution that partially captures the true tumor -- a sign of genuine tissue learning within the limits of available training data.
The cascaded randomization test revealed that even with progressively randomized network weights, some structure from the input image persisted in the heat maps -- with MI and SSIM values settling at approximately 0.26 and 0.22 rather than 0. The authors attribute this to the U-Net's skip connections, which pass spatial information from the encoder to the decoder regardless of learned weights. This identifies a systematic limitation of applying Grad-CAM to U-Net architectures that future explainability methods will need to address.
The study also highlights that histopathology-based validation introduces its own challenges: only patients who underwent prostatectomy have resected tissue available, biasing the validation set toward intermediate and high-risk cancers. Additionally, prostate deformation during specimen processing introduces co-registration errors that propagate into the histopathology-MRI alignment, potentially degrading the quality of the gold standard itself.
This study demonstrates that combining whole-mount histopathology validation with Grad-CAM explainability provides a more complete evaluation of prostate cancer AI segmentation models than comparison with radiologist contours alone. The CNN passed all interpretability sanity checks and its predictions agreed comparably with both the clinical (MRI) and pathological (histology) ground truths -- a meaningful finding given that the network was trained only on MRI-based labels.
The results support using such a CNN as a tool for generating an initial tumor contour that a radiation oncologist or radiologist can then review and refine, rather than as a fully autonomous system. This human-in-the-loop approach is appropriate given the current performance levels and the high clinical stakes of prostate tumor delineation for radiation treatment planning.
Future improvements should focus on training with higher-quality labels, potentially derived from co-registered histopathology in larger patient cohorts. As MRI-histopathology co-registration techniques improve in accuracy and are applied to larger datasets, CNNs trained on these annotations are expected to achieve substantially better segmentation performance and more reliable explainability outputs than currently achievable.