Ai-Based Grading Approach Identifies Fnclcc Grade 3 Soft Tissue Sarcomas

Morphologic grading is an established prognostic parameter in soft tissue sarcomas. Three-tiered FNCLCC grading is highly correlated with overall survival of sarcoma patients and is therefore widely used for deciding on perioperative chemo- and radiotherapy. Clinical routine, however, faces two major limitations of conventional FNCLCC grading: a) Grading correctness is hampered by intratumoral heterogeneity, especially in small biopsies; b) G2 sarcomas represent a clinically equivocal subgroup but account for the majority of patients. We selected H&E-stained sections of 52 soft tissue sarcomas. At least 13 mm2 of viable tumor regions were annotated (necroses were excluded) per case. After scanning the resulting 240 slides at 20 × resolution, we extracted approximately 1.3 million image tiles (112.6×112.6 μm) of the tumor regions. We trained a convolutional neural network (customized pre-trained EfficientNetB0) to classify tiles in high grade (FNCLCC G3; 24 cases) and non-high grade (G1 or G2; 28 cases) by predicting the respective class probabilities. For training, we used 41 cases (incl. 11 for validation), balancing the numbers of tiles per case, and withheld 11 test cases for independent evaluation. Finally, we optimized the cut-off values of a) the class probability for classifying tiles and b) the proportion of high-grade tiles required for classifying a case as high-grade. For default cut-off values of 0.5, we obtained accuracies of 68% (slide-based) and 64% (case-based) for the test cases. Optimizing the cut-off values, accuracies increased to 77% (slide-based) and 73% (case-based). G3 tumors were correctly classified as high grade (3/4), except for one pleomorphic liposarcoma, probably misclassified due to specific tumor morphology. G2 tumors were classified either as low grade (4/6) or high grade (2/6), one G1 tumor was classified correctly. Our data indicate that computerized grading is feasible for sarcomas and AI-based analysis can recognize G3 sarcomas correctly, without being predicted on heterogeneous parameters, e.g., mitotic count and extent of necroses. Possibly observable patterns in the tile-wise classification of G2 sarcomas require further investigation and correlation with clinical outcome data.