Artificial intelligence significantly improves the diagnostic accuracy of deep myxoid soft tissue lesions in histology

Abstract

Deep myxoid soft tissue lesions have posed a diagnostic challenge for pathologists due to significant histological overlap and regional heterogeneity, especially when dealing with small biopsies which have profoundly low accuracy. However, accurate diagnosis is important owing to difference in biological behaviors and response to adjuvant therapy, that will guide the extent of surgery and the need for neo-adjuvant therapy. Herein, we trained two convolutional neural network models based on a total of 149,130 images representing diagnoses of extra skeletal myxoid chondrosarcoma, intramuscular myxoma, low-grade fibromyxoid sarcoma, myxofibrosarcoma and myxoid liposarcoma. Both AI models outperformed all the pathologists, with a significant improvement of accuracy up to 97% compared to average pathologists of 69.7% (p < 0.00001), corresponding to 90% reduction in error rate. The area under curve of the best AI model was on average 0.9976. It could assist pathologists in clinical practice for accurate diagnosis of deep soft tissue myxoid lesions, and guide clinicians for precise and optimal treatment for patients.

Figure 1

Confusion matrix of pathologists (A), DenseNet-121 (B) and EfficientNet B3 (C). Both deep learning models have significant improvement in accuracy compared to pathologists (D), and EfficientNet B3 has a small significant further improvement compared to DenseNet-121. EMC extraskeletal myxoid chondrosarcoma, IM intramuscular myxoma, LGFMS low grade fibromyxoid sarcoma, MFS myxofibrosarcoma, MLS myxoid liposarcoma.

Figure 2

Overview of strategies of training the deep learning models in comparison with pathologists (A). Number of images in different diagnostic categories of different datasets (B). Schematic diagram depicting the training workflow of AI model (C). EMC extraskeletal myxoid chondrosarcoma, IM intramuscular myxoma, LGFMS low grade fibromyxoid sarcoma, MFS myxofibrosarcoma, MLS myxoid liposarcoma.

Figure 3

The precision, recall and F1-score for different diagnostic categories of pathologists (A) and the deep learning models (B). EMC extraskeletal myxoid chondrosarcoma, IM intramuscular myxoma, LGFMS low grade fibromyxoid sarcoma, MFS myxofibrosarcoma, MLS myxoid liposarcoma.

Figure 4

Receiver operating characteristic curve and Precision-Recall curve of DenseNet-121 (A) and EfficientNet B3 (B). EMC extraskeletal myxoid chondrosarcoma, IM intramuscular myxoma, LGFMS low grade fibromyxoid sarcoma, MFS myxofibrosarcoma, MLS myxoid liposarcoma.

Figure 5

Example images and heatmaps generated by the deep learning model of (A,B) Intramuscular myxoma, (C,D) Myxoid liposarcoma, (E–H) Myxofibrosarcoma, (I,J) Extraskeletal myxoid chondrosarcoma, and (K,L) Low-grade fibromyxoid sarcoma. The images are mainly activated over tumour nuclei and the vessels. There are subtle sub-visual differences of tumour nuclei and vessels, especially between intramuscular myxoma and myxofibrosarcoma, that are not noticeable by human eyes.