Computational analysis of pathological images enables a better diagnosis of TFE3 Xp11.2 translocation renal cell carcinoma

Abstract

TFE3 Xp11.2 translocation renal cell carcinoma (TFE3-RCC) generally progresses more aggressively compared with other RCC subtypes, but it is challenging to diagnose TFE3-RCC by traditional visual inspection of pathological images. In this study, we collect hematoxylin and eosin- stained histopathology whole-slide images of 74 TFE3-RCC cases (the largest cohort to date) and 74 clear cell RCC cases (ccRCC, the most common RCC subtype) with matched gender and tumor grade. An automatic computational pipeline is implemented to extract image features. Comparative study identifies 52 image features with significant differences between TFE3-RCC and ccRCC. Machine learning models are built to distinguish TFE3-RCC from ccRCC. Tests of the classification models on an external validation set reveal high accuracy with areas under ROC curve ranging from 0.842 to 0.894. Our results suggest that automatically derived image features can capture subtle morphological differences between TFE3-RCC and ccRCC and contribute to a potential guideline for TFE3-RCC diagnosis.

Fig. 1. Workflow scheme.

a H&E-stained tissue slides were digitized by a scanner to obtain whole-slide images. b A large set of quantitative image features were extracted, characterizing nucleus size, staining, shape, and density. c The Mann–Whitney U test was used to compare image features between TFE3-RCC and ccRCC, and machine learning models were developed based on the image features to automatically classify the two cancer subtypes. On the box plots in c, the central mark indicates the median, and the bottom and top edges of the box indicate the 25th and 75th percentiles (q₁ and q₃), respectively. The upper whisker extends from q₃ to q₃ + 1.5 × (q₃ − q₁), and the lower whisker extends from q₁ to q₁ − 1.5 × (q₃ − q₁), while data beyond the end of the whiskers are outlying points that are plotted individually.

Fig. 2. Feature extraction pipeline.

a The nuclei in whole-slide images are automatically segmented. b For each segmented nucleus, 10 nucleus-level features, regarding nucleus size, staining intensity, shape, and density, are extracted. c For each type of nucleus-level features from the same whole-slide image, they are dissected into 15 image-level features using a 10-bin histogram and five distribution statistics. Scale bars: 5 mm (a whole-slide image) and 50 µm (a image patch).

Fig. 3. Comparison of image features between TFE3-RCC and ccRCC.

For each feature, the fold change is defined as the ratio of the median feature values between ccRCC and TFE3-RCC. 52 image features that show significant differences between TFE3-RCC and ccRCC are identified using the two-sided Mann–Whitney U test. Multiple comparison correction is performed using false discovery rate procedure at 5% level.

Fig. 4. The performance of the four machine learning models.

a Classification performance on dataset 1 using five-fold cross-validation (n = 5 experiments for each model). For each, 80% of patients were used as the training set and the remaining patients were used as the internal validation set. b Receiver operating characteristics curves for classifying TFE3-RCC and ccRCC in the external validation set (dataset 2). Models were trained using dataset 1 and evaluated using dataset 2. The 95% confidence intervals for the AUC: LR (0.763–0.984), RF (0.736–0.960), SVM-L (0.725–0.959), and SVM-G (0.797–0.991). LR, logistic regression; RF, random forest; SVM-L, SVM with linear kernel; SVM-G, SVM with Gaussian kernel. Data are represented as mean ± SD in a.