A consequential diagnostic challenge. Bladder cancer staging at early non-muscle invasive disease determines treatment intensity and long-term prognosis. Stage Ta tumors are non-invasive papillary lesions confined to the urothelium and are typically manageable conservatively, while T1 tumors have invaded the subepithelial connective tissue and carry a substantially higher risk of progression to muscle-invasive and potentially fatal disease.
Despite these very different prognoses, Ta and T1 tumors can appear remarkably similar under the microscope. In one landmark restaging study, 235 tumors initially classified as T1 were reclassified in 35% of cases as Ta when reviewed by an experienced pathologist, illustrating the degree of inter-observer disagreement that exists even among experts.
An urgent need for objective tools. No molecular biomarker currently distinguishes Ta from T1 tumors reliably. Histological assessment remains the standard approach, but its subjective nature and dependence on pathologist experience introduce unacceptable variability in staging decisions. An objective, machine learning-based system that quantifies the relevant microscopic patterns could substantially improve staging consistency and clinical outcomes.
Dataset and known pathological patterns. A total of 1,177 H&E stained histopathology images were collected from the University of Rochester Medical Center, comprising 460 non-invasive Ta and 717 invasive T1 tumors. Images were captured at 100x magnification as 2048 by 2048 pixel files and tiled to 700 by 700 pixel regions. Only cases that genitourinary pathologists could confidently stage were included, providing clean ground-truth labels.
The feature extraction strategy was built around three pathologist-identified microscopic patterns that distinguish T1 from Ta tumors: desmoplastic reaction, which is dense fibrosis surrounding tumor cell nests; retraction artifact, the tissue shrinkage gap seen around invading tumor nests after tissue processing; and abundant pinker cytoplasm in T1 cells due to higher eosin uptake. Separate automated pipelines were built for each pattern using ImageJ and CellProfiler.
Nearly 700 quantitative features extracted. ImageJ extracted 60 textural features across the whole tissue including crack edge area, pixel intensity distributions in connective tissue and cytoplasm, while CellProfiler extracted 636 cell-level features including nuclear size distributions, nuclear shape descriptors, and connective tissue shape. Combined, nearly 700 features per image were extracted, providing a rich quantitative representation of the three invasion patterns.
The retraction artifact pipeline illustrates the specificity of the approach: color thresholding separated tissue from non-tissue, diameter-based filtering isolated small pericellular cracks from large inter-tissue spaces, and pixel count in the resulting mask quantified the extent of retraction artifact. This type of domain knowledge-guided pipeline enabled precise capture of biologically meaningful features that generic image analysis software would miss.
Unsupervised clustering fails. Dimensionality reduction via principal component analysis followed by k-means clustering with k values from 2 to 9 produced no recognizable separation between Ta and T1 images. Information gain from any cluster split was less than 0.006, confirming that the non-invasive and invasive images are not linearly separable and that supervised learning is necessary.
After filtering out non-informative features, 696 features were retained and ranked by decision tree importance. Performance across six classifiers increased as the top-ranked features were added and plateaued between the 70th and 100th features, after which adding more features degraded performance. The top 100 features outperformed the full 696-feature set in five of six classifiers, demonstrating the value of principled feature selection in avoiding overfitting.
High accuracy across six ML methods. Using the top 100 features, all six classifiers, namely probabilistic neural network, random forest, support vector machine, bagging, logistic regression, and multilayer perceptron, achieved greater than 90% accuracy. The probabilistic neural network achieved the best performance with an accuracy of 96.7% and an AUC of 0.99, significantly exceeding the CNN-based approach described next.
CNN accuracy capped at 84%. Pre-trained VGG16 and VGG19 convolutional neural networks, which automatically extract features from images, achieved accuracies of 84% and 81% respectively, with AUC values of 0.926 and 0.912. While these are respectable results, they fell 10 to 15 percentage points below the top supervised ML classifiers using domain-knowledge-guided features.
The finding directly challenges a common assumption in computational pathology: that end-to-end deep learning automatically extracts better features than human-guided feature engineering. For this specific and challenging classification task, encoding pathologist knowledge into the feature extraction process provided a meaningful performance advantage.
Why domain knowledge helps here. The VGG networks used were pre-trained on general photographic images rather than histopathological slides, which may limit their ability to detect the subtle tissue-level patterns that distinguish Ta from T1 disease. Furthermore, with only around 1,000 images, the dataset may be too small to allow a CNN to learn these domain-specific features from scratch. The authors suggest that histopathology-specific pre-training and CNN intermediate layer analysis could close part of the performance gap in future work.
Pattern-specific performance ranking. Separating the features into the three invasion patterns revealed clear differences in their individual predictive power. The desmoplastic reaction pattern, which had 675 features, achieved 90.5% accuracy with an average AUC of 0.98 across six classifiers. Pinker cytoplasm features, represented by only 13 features, achieved 74.5% accuracy with an AUC of 0.825. Retraction artifact, with 15 features, achieved 73.4% accuracy and an AUC of 0.802.
Even after reducing the desmoplastic reaction feature set from 675 to 70 features, performance still exceeded 90% and remained higher than either of the other two patterns, confirming that the superior performance of desmoplastic reaction features is not simply due to having more features available. This pattern contains the most diagnostically informative signal for distinguishing T1 invasion.
Nuclear size and count are the most predictive individual features. Within the desmoplastic reaction pattern, the top-ranked individual features were the number of nuclei and the distribution of nuclei sizes. This is biologically consistent: desmoplastic reaction is characterized by dense inflammatory cell infiltration surrounding tumor nests, creating a distinctive high-cellularity microenvironment with many small inflammatory nuclei adjacent to larger tumor cell nuclei that produces a quantitatively distinctive nuclear landscape.
An interpretable diagnostic aid. Unlike black-box deep learning models, the feature-centered approach produces probability scores tied to specific biologically meaningful patterns such as the extent of desmoplastic reaction or nuclear characteristics. This interpretability is essential for clinical adoption, as pathologists need to understand why the system made a particular prediction to incorporate it into their decision-making process.
The authors distinguish between interpretability and causability, arguing that reaching the level where a model explains the causal mechanism of invasion, not just predicts the label, is the ultimate clinical goal. Achieving causability would allow pathologists to use the system not only as a second opinion but as a tool for exploring and teaching the pathological basis of T1 invasion.
Potential for broader application. The pipeline architecture, in which domain experts define the patterns of interest and automated image analysis tools extract quantitative features from those patterns, is generalizable to other cancer types where specific histological invasion patterns have been characterized. Combining features identified from CNN intermediate layers with the expert-guided features is a promising direction for further improving accuracy while maintaining interpretability.