Matteo Sacco1, Erica Pietroluongo1,2, Aliya N. Husain3, Qudsia Arif3, Alessandra Esposito1, Anna Di Lello1, Marina Chiara Garassino1, Mirella Marino4, James M. Dolezal5
1Department of Medicine, Section of Hematology/Oncology, University of Chicago, Chicago, IL, USA;
2Department of Clinical Medicine and Surgery, University of Naples Federico II, Naples, Italy;
3Department of Pathology, University of Chicago, Chicago, IL, USA;
4Department of Pathology, IRCCS Regina Elena National Cancer Institute, Rome, Italy;
5Geisinger Cancer Institute, Danville, PA, USA
Correspondence to: Matteo Sacco, MS. Department of Medicine, Section of Hematology/Oncology, University of Chicago, 900 E. 57th St., Chicago, IL 60637, USA. Email: mattsacco@bsd.uchicago.edu.
Background: Thymic epithelial tumors (TETs) pose significant diagnostic challenges due to their diverse histology, which can impact treatment decisions. Accurate classification of TET subtypes is crucial, and second pathological opinions often result in changes to diagnosis and treatment plans. Recent advancements have leveraged artificial intelligence (AI) to enhance pathology diagnostic accuracy. Our previous study, utilizing a deep convolutional neural network, analyzed digital hematoxylin and eosin (H&E) slides from The Cancer Genome Atlas (TCGA) to classify TET subtypes. The findings demonstrated that the AI model effectively distinguished between A/AB, B1 and B2/B3 thymoma subtypes, highlighting the potential of AI to improve diagnostic consistency for these complex tumors.
Methods: We used the same model architecture (Xception-based deep convolutional neural network model), trained it on all digital H&E stained slides from TCGA to predict histologic subtypes of thymomas (A, AB, B1, B2, and B3) as an ordinal variable. The model was then tested on an external validation set comprised of 88 slides from UChicago. Model predictions were compared between groups using one-way analysis of variance (ANOVA), pairwise t-tests with Bonferroni correction for multiple comparisons. We also calculated the R-squared value to assess the model’s predictive performance.
Results: The deep learning predictions among the TET subtypes were significantly different by ANOVA (F-statistic =13.92, P<0.001). Pairwise comparisons revealed significant differences between several subtype pairs, notably between B1 and B2 (P<0.001), and B2 and B3 (P=0.04). The model showed moderate predictive performance with an R-squared value of 0.38.
Conclusions: This external validation study confirms that deep learning can discriminate between TET histologic subtypes using digital H&E slides. The model demonstrates strength in distinguishing B2 thymomas from other subtypes, a novel finding not observed in our previous study. While the overall predictive performance is moderate, these results support the potential of AI-assisted diagnosis in rare tumors like TETs. Given these initial results, our aim is to increase the training sample size along with optimizing the model architecture to improve discriminative power.
Keywords: Thymic epithelial tumors (TETs); deep learning; computational pathology
