Purpose: To develop an end-to-end deep learning (DL) model for the diagnosis of benign and malignant pelvic and sacral tumors (PSTs) using non-enhanced magnetic resonance imaging (MRI).

Materials and methods: A total of 835 patients with PSTs at four hospitals were retrospectively used for model training, validation and testing. We built six diagnostic models based on original images, radiologist’s label, segmentation model with or without clinical data, segmentation model based on non-enhanced MRI with or without clinical data, respectively. The diagnosis and reading time of three radiologists were recorded for comparison with the models. The models were assessed using the area under the curve (AUC) and accuracy (ACC) values.

Results: Our proposed Model SEG-CL-NC achieved an AUC of 0.823, and 0.776 ACC for diagnosis in the Internal Test Set 1, and 0.836 AUC, and 0.781 ACC in the Internal Test Set 2. The ACC of External Dataset Centers 2, 3, and 4 were 0.714, 0.740, and 0.756, and the sensitivity was 0.690, 0.660, and 0.704, and the specificity was 0.762, 0.917, and 0.857, respectively. The Model SEG-CL-NC’s ACC was comparable to that of enhanced model and radiologists (P > 0.05). However, the diagnosing time of Model SEG-CL-NC is significantly shorter than radiologists (P < 0.01).

Conclusion: Our results suggested that the proposed end-to-end DL model using non-enhanced MRI could achieve comparable performance to contrast-enhanced models and radiologists in diagnosing benign and malignant PSTs, highlighting its potential as an accurate, efficient, and cost-effective tool for clinical practice.