Development and External Validation of Deep-Learning-Based Tumor Grading Models in Soft-Tissue Sarcoma Patients Using MR Imaging

SIMPLE SUMMARY: In soft-tissue sarcoma (STS) patients, the decision for the optimal treatment modality largely depends on STS size, location, and a pathological measure that assesses tumor aggressiveness called “tumor grading”. To determine tumor grading, invasive biopsies are needed before therapy. In previous research studies, quantitative imaging features (“radiomics”) have been associated with tumor grading. In this work, we assessed the possibility of predicting tumor grading using an artificial intelligence technique called “deep learning” or “convolutional neural networks”. By analyzing either T1-weighted or T2-weighted MRI sequences, non-invasive tumor grading prediction was possible in an independent test patient cohort. The results were comparable to previous research work obtained with radiomics; however, the reproducibility of the contrast-enhanced T1-weighted sequence was improved. The T2-based model was also able to significantly identify patients with a high risk for death after therapy. ABSTRACT: Background: In patients with soft-tissue sarcomas, tumor grading constitutes a decisive factor to determine the best treatment decision. Tumor grading is obtained by pathological work-up after focal biopsies. Deep learning (DL)-based imaging analysis may pose an alternative way to characterize STS tissue. In this work, we sought to non-invasively differentiate tumor grading into low-grade (G1) and high-grade (G2/G3) STS using DL techniques based on MR-imaging. Methods: Contrast-enhanced T1-weighted fat-saturated (T1FSGd) MRI sequences and fat-saturated T2-weighted (T2FS) sequences were collected from two independent retrospective cohorts (training: 148 patients, testing: 158 patients). Tumor grading was determined following the French Federation of Cancer Centers Sarcoma Group in pre-therapeutic biopsies. DL models were developed using transfer learning based on the DenseNet 161 architecture. Results: The T1FSGd and T2FS-based DL models achieved area under the receiver operator characteristic curve (AUC) values of 0.75 and 0.76 on the test cohort, respectively. T1FSGd achieved the best F1-score of all models (0.90). The T2FS-based DL model was able to significantly risk-stratify for overall survival. Attention maps revealed relevant features within the tumor volume and in border regions. Conclusions: MRI-based DL models are capable of predicting tumor grading with good reproducibility in external validation.