EP093/#664 Automatic multimodal classification using transformer for cervical cancer

Objectives

In cervical cancer diagnosis, automatic classification and identification can effectively reduce the workload of radiologists and gynecologists. Due to that T2-MR and DWI-MR images are complementary in lesion information, it is necessary to combine medical images with the two modalities. In this study, we proposed a automatic classification pipeline for cervical cancer based on Swin-Transformer and verified the classification potential on multimodality.

Methods

Fitty-eight T2-MR images and eighty DWI-MR images of patients with cervical cancer were retrospectively enrolled. Totally 1858 slices were annotated by radiologists to four classes, including T2-tumor, T2-notumor, DWI-tumor and DWI-notumor. 1489 slices were used for training and 369 images for validation. In addition, images of ten patients containing 184 DWI slices and 200 T2 slices were not participated in modeling as test set. All the gray slices(512×512) with single channel were repeated to three channels and resized to 224×224 pixels. Finally, the mixed 1858 slices(892 DWI, 966 T2) were put into the classification network based on Swin-Transformer (optimizer: AdamW, batchsize = 8, lr = 0.0001).

Results

Four typical metrics were applied to evaluate the classification result, including accuracy, F1-score, sensitivity and specificity. Regarding to the 184 DWI test slices, the accuracy, F1-score, sensitivity, specificity were 91.85%, 90.15%, 88.52% and 93.50%, respectively. And the accuracy of 200 T2 test slices was 83.00% . Besides, there was no error in the distinguishment between the two modalities.

Conclusions

This work confirmed that the tumor of cervical cancer on multimodal MRI images can be automatically classified with high accuracy.