Deep learning-based histopathological subtyping of peripheral T-cell lymphoma (PTCL) on H&E-stained whole slide images

Abstract

Abstract Introduction Peripheral T-cell lymphomas (PTCL) are rare and highly heterogenous malignancies. The diagnosis can be challenging for pathologists. Accurate subtyping of PTCL is crucial for therapy decision making and clinical outcome prediction. Here we demonstrate that a deep learning (DL) model can reliably identify common subgroups of PTCL using histopathological images. Methods We trained a DL model on 223 H&E stained whole slides images (WSIs) from the German National T-Cell Lymphoma Registry and Biobank of the GLA and OSHO study groups. The training cohort included 144 patients with the following subtypes: anaplastic large cell T-cell lymphoma (ALCL) n=31; peripheral T-cell lymphoma with T follicular-helper phenotype (TFH-NHL) n=46; peripheral T-cell lymphoma, not otherwise specified (PTCL-NOS) n=23; other subtypes n=44. We trained a weakly supervised, open-source DL pipeline comprising UNI model as feature extractor and a vision transformer architecture for multiclass prediction. The DL model was trained to differentiate between the four subgroups: ALCL, TFH-NHL, PTCL-NOS and a group of other T-NHL entities. External validation was performed using an independent cohort of 118 WSIs from 118 patients (ALCL n=25; TFH-NHL n=75; PTCL-NOS n=18). To assess the quality of the model we used area under the receiver operating characteristic curve (AUROC). Heatmaps highlighting the most important image areas leading to correctly and incorrectly predicted cases were evaluated by hematopathologists to identify patterns for incorrectly and correctly categorized cases. Results In the five-fold cross-validation the DL model significantly distinguished the subgroups with an AUROC of 0.9 (95% confidence interval 0.84-0.96) for ALCL, TFH-NHL with 0.77 (0.59-0.95), PTCL-NOS with 0.55 (0.41-0.69) and other subtypes with 0.81 (0.71-0.90), even in different tissues types. When applied to the external validation cohort, the model maintained high predictive performance with AUROCs of 0.85 (0.74-0.95) for ALCL, 0.87 (0.79-0.94) for PTCL-TFH and 0.69 (0.52-0.85) for PTCL-NOS. A comparison of the most important image areas for DL-based classification was carried out with reference pathologists. This showed that the DL model successfully recognized the characteristic morphological areas, resulting in highly accurate classification for ALCL and PTCL-TFH. Conclusions The DL model was able to classify main PTCL subtypes on H&E stainings with high accuracy. This model needs prospective validation as a supportive tool for PTCL diagnosis. HE-based DL-based subtyping could assist pathologists in decision-making and help to guide effective diagnostics, addressing a critical unmet need.