Predicting ADC Map Quality from T2-Weighted MRI: A Deep Learning Approach for Early Quality Assessment to Assist Point-of-Care.

Purpose: Poor quality prostate MRI images, especially ADC maps, can lead to missed lesions and unnecessary repeat scans. To address this issue, we aimed to develop an automated method to predict ADC map quality from T2 images acquired earlier in the scanning process. Materials and Methods: A paired multi-site image corpus of T2-weighted images and ADC maps was constructed from 486 patients imaged in-house and at 62 external clinics. A senior radiologist assigned 1-3 quality ratings to each image set, later converted to a binary "non-diagnostic" or "diagnostic" scale. A deep learning model and a rectal cross-sectional area measurement approach were developed to predict ADC image quality from T2 images. Model performance was evaluated retrospectively by accuracy, sensitivity, negative and positive predictive value, and AUC. Results: No single acquisition parameter in the metadata was statistically associated with image quality for either T2 or ADC maps. Quality scores of the same modality showed low correlation across sites (r~0.2). In the challenging task of predicting ADC quality from prior T2 images, our model achieved performance comparable to current single-site models directly using ADC maps, with 83% sensitivity and 90% negative predictive value. The model showed stronger performance on in-house data (94±2% accuracy) despite being trained exclusively on multicenter external data. Rectal cross-sectional area on T2 images provided an interpretable quality metric (AUC 0.65). Conclusion: The probability of low quality, uninterpretable ADC maps can be inferred early in the imaging process by neural network approach, allowing corrective action to be employed.