Objective To explore the feasibility of diagnosing diabetic foot (DF) based on radiomic features of magnetic resonance imaging (MRI) .

Methods The clinical data of 127 patients with foot diseases who underwent foot MRI examination in Shanghai University of TCM, Shanghai TCM-integrated Hospital from August 2018 to August 2020 were analyzed retrospectively, including 83 males and 44 females, aged 16–88 (59.3±14.8) years. The patients were divided into the DF group (85 cases) and non-DF group (42 cases) according to the clinical diagnosis of the patient. Based on the features of radiomics, using simple random sampling method, the patients were divided into the training group (76 cases) and test group (51 cases) at the ratio of 3∶2. First, the cuboid was sketched on sagittal images of the T1-weighted imaging (T1WI) and proton density weighted imaging (PDWI) fat suppression sequence of MRI, and the radiomic features were extracted. Then, the T1WI sequence, PDWI fat suppression sequence, and combined models were constructed; the combined model included gender, age, bone marrow signals, and radiomic features. Based on the pretreatment of normalization of the maximum absolute value of three groups of model parameters, the optimal feature dimensions were selected using optimal feature and model selection. The age between the two groups was compared using the independent-sample t test, and the gender and abnormal bone marrow signal of midfoot between the two groups were compared using the chi-square test.

Results Significant differences in age, gender ratio, and abnormal bone marrow signal of midfoot were observed between the DF group and non-DF group (t=29.2; χ²=17.15, 6.53; all P<0.05). The T1WI sequence model, PDWI fat suppression sequence model, and combined model finally screened 9, 7, and 10 optimal feature dimensions, respectively. The support vector machine model was used to distinguish DF and non-DF patients. On the T1WI sequence model, the area under the curve (AUC) of the training and test groups were 0.86 and 0.84, respectively, and the accuracy rates were 78% and 69%, respectively. On the PDWI fat suppression sequence model, the AUC of the training and test groups were 0.85 and 0.83, respectively, and the accuracy rates were 82% and 78%, respectively. On the combined model, the AUC of the training and test groups were 0.93 and 0.85, respectively, and the accuracy rates were 84% and 76%, respectively. The diagnostic efficacy of the T1WI sequence model was equivalent to that of the PDWI fat suppression sequence model, whereas the combined model showed better diagnostic efficacy than the two aforementioned models.

Conclusion MRI radiomics can effectively distinguish DF from non-DF, which can provide a new, efficient, and noninvasive method for DF imaging diagnosis.