Multiparametric MRI (mpMRI) has become the standard imaging approach for evaluating prostate cancer, combining structural T2-weighted imaging with functional diffusion and contrast sequences. Its interpretation is standardized through the PI-RADS scoring system, which rates lesions from 1 (very unlikely cancer) to 5 (highly likely cancer).
Despite its widespread adoption, PI-RADS has a significant limitation: it requires a high level of radiological expertise and exhibits substantial variability between readers. Two radiologists examining the same scan can assign different PI-RADS scores, meaning patients may receive different clinical recommendations depending on who reads their scan.
The problem is particularly acute for PI-RADS 3 lesions, which are classified as equivocal. Studies show that only 15 to 35% of PI-RADS 3 lesions are actually biopsy-positive for clinically significant cancer, yet current practice frequently sends these patients for biopsy anyway, leading to a large number of unnecessary invasive procedures.
An automated system that uses quantitative image features to refine the classification of suspicious lesions beyond what PI-RADS scoring alone can provide could reduce overdiagnosis and unnecessary biopsies while ensuring that significant cancers are not missed.
Researchers at UCLA designed a Textured Deep Learning (Textured-DL) model that combines classical texture analysis with convolutional neural networks to classify prostate lesions already identified on mpMRI as either clinically significant prostate cancer (Gleason Score 3+4 or above) or non-clinically significant (Gleason Score 3+3 or lower, or false positives).
The core innovation is using a 3D Gray-Level Co-occurrence Matrix (GLCM) extractor as the first step, converting the raw MRI image patch around a lesion into a mathematical representation of its texture before passing that representation to a neural network for classification.
This approach was designed to overcome a limitation of standard deep learning applied to MRI: when lesion patches are resized to a fixed dimension before entering the network, scale information about the tumor is lost. By working with texture matrices derived from the original volumetric patch, Textured-DL preserves the spatial relationships that encode tumor heterogeneity.
The study enrolled 402 patients from UCLA who underwent 3T mpMRI followed by robotic-assisted radical prostatectomy, with whole-mount histopathology providing lesion-by-lesion ground truth. This large dataset with surgical pathology correlation is substantially more rigorous than biopsy-based labeling.
For each suspicious lesion identified on mpMRI, a 3D volumetric patch from both the T2-weighted image and the ADC map is cropped to tightly surround the lesion boundary. These patches are intensity-normalized to a 0-255 range, then processed by the GLCM extractor.
The 3D GLCM counts how frequently pairs of voxels at specific spatial distances and orientations within the lesion share particular intensity values. Unlike the 2D GLCM used in conventional radiomics, the 3D version considers voxel relationships in all three spatial dimensions, capturing through-plane texture patterns that are lost in slice-by-slice analysis.
For each lesion patch, 13 separate GLCMs are computed corresponding to 13 spatial directions around each voxel. This produces a rich 3D representation of how the intensity pattern is organized throughout the tumor volume, quantifying properties like uniformity, complexity, and directional periodicity of the tissue texture.
The T2-weighted and ADC GLCM sets are then concatenated and fed into a compact convolutional neural network with two convolutional layers and two fully connected layers. The CNN learns which texture patterns in which combinations are most predictive of clinically significant cancer, outputting a single probability score for each lesion.
Testing on 121 held-out patients containing 142 lesions, Textured-DL achieved an AUC of 0.85 for distinguishing clinically significant from non-clinically significant prostate cancer. This was significantly higher than PI-RADS-based classification at a cutoff of 4 or above, which achieved an AUC of 0.73 (p less than 0.05).
The specificity difference was especially striking: Textured-DL achieved a specificity of 0.70 compared to only 0.47 for PI-RADS classification, a statistically significant difference. This means Textured-DL correctly identified as benign a much higher proportion of non-cancerous lesions, which is the performance attribute most directly related to reducing unnecessary biopsies.
Textured-DL also outperformed all other machine learning baselines tested in the study: a random forest classifier using handcrafted texture features (AUC 0.78), a plain image-based CNN (AUC 0.74), a deep CNN with transfer learning (AUC 0.76), and a deeper VGG-style CNN (AUC 0.76). These comparisons confirm that combining texture matrices with deep learning outperforms both feature engineering and raw image deep learning separately.
The advantage of Textured-DL over standard CNNs is partly explained by the fact that standard approaches must resize lesion patches to a fixed input size, destroying scale information. Clinically, tumor size is meaningful: smaller lesions are more likely to be indolent. Texture-based representation preserves this information within the statistical properties of the GLCM.
In the peripheral zone (PZ), where 80% of prostate cancers originate, Textured-DL achieved an even higher AUC of 0.88, with a specificity of 0.78 compared to only 0.42 for PI-RADS classification (p less than 0.05). This large specificity gap suggests that PI-RADS struggles particularly in the peripheral zone, where the cancer-related signal changes can be subtle against normal peripheral zone tissue.
In the transition zone (TZ), where benign prostatic hyperplasia creates confounding background changes, Textured-DL performed similarly to all baseline methods, with no statistically significant advantages. This makes biological sense: TZ cancer detection is inherently more challenging due to the complex tissue background, and the texture-based features may not fully separate cancer from hyperplastic tissue.
For solitary tumors (single lesion), Textured-DL showed a significantly higher AUC than most baselines and a much higher specificity than PI-RADS classification, suggesting the approach works best when there is a single dominant lesion to characterize. For multi-focal tumors (multiple lesions), the AUC advantage was less consistent, possibly because complex interactions between multiple lesions add classification difficulty.
These zone-specific and focality-specific results reinforce a broader principle in prostate cancer AI: the optimal algorithm and feature set depends significantly on where in the gland the tumor is located, and blanket approaches applied uniformly across all prostate zones will always leave performance gains on the table.
Among the 43 PI-RADS 3 lesions in the test set, where only 13 (30%) were clinically significant cancer, Textured-DL achieved an AUC of 0.81 and correctly classified 85% of the actual cancers (sensitivity) and 73% of the non-cancerous lesions (specificity). This is a clinically meaningful result given that this is precisely the category where current practice is most uncertain.
For PI-RADS 3 lesions, the negative predictive value (NPV) of Textured-DL was 92%, meaning that when the model predicts a PI-RADS 3 lesion is non-cancerous, it is correct 92% of the time. The positive predictive value was 58%. In comparison, PSA density (PSAD) thresholds commonly used in practice achieved NPVs of only 67 to 75%.
This high NPV is the clinically most important finding for the PI-RADS 3 population: it means the model can reliably identify which equivocal lesions are very likely to be non-cancerous, enabling a physician to safely defer biopsy for those patients. The approximately one in twelve patients who would be incorrectly reassured represents an acceptable trade-off if the alternative is biopsying everyone.
The authors propose that Textured-DL could function as a second-reader decision support tool, applied after radiologists identify and contour PI-RADS 3 lesions, to provide a quantitative probability score that supplements clinical judgment in the biopsy decision.
The study was conducted at a single academic center with highly experienced genitourinary radiologists, meaning PI-RADS performance was at its best in this dataset. In community settings with less experienced readers, the absolute advantage of Textured-DL might be even larger, but the model's own performance would need validation in those environments.
The patient cohort consisted entirely of men who underwent radical prostatectomy, meaning they already had known or suspected significant cancer. This creates a higher baseline cancer prevalence than would be seen in a screening or biopsy-planning population, which inflates the positive predictive value and may affect generalizability to patients with lower pre-test probability.
The current model uses only T2-weighted and ADC images. The authors note that adding high b-value DWI, DCE MRI, and potentially oxygen-enhanced MRI sequences could further improve performance, as each sequence provides complementary biological information. Incorporating clinical variables like PSA density, patient age, and lesion location could also improve the model.
Multi-center validation with diverse scanners and reader populations remains the critical next step before clinical deployment. The patterns of texture features that predict cancer at UCLA may not perfectly generalize to institutions with different scanner types, imaging protocols, or patient demographics, requiring either a universal model or institutional fine-tuning.