Ischemic stroke, a major global health burden, requires accurate collateral circulation assessment for optimal diagnosis, treatment, and prognosis. The status of collateral circulation can be accurately assessed using techniques such as CT perfusion imaging (CTP), CT angiography (CTA), and multiphase CT angiography (mCTA). Nevertheless, achieving an accurate assessment of a patient’s collateral circulation status requires extensive clinical knowledge and expertise in interpreting medical images. In such scenarios, leveraging deep learning technology to aid in medical diagnosis can boost diagnostic efficiency, facilitate more precise and effective collateral circulation assessment, and ultimately enhance the prognosis of stroke patients. In this study, we constructed a mCTA imaging dataset and proposed a convolutional neural network mode incorporating the Hybrid Multiple Phase Attention (HMPA) module, termed the mCTAformer network, for collateral circulation assessment. The HMPA module not only focuses on features from different time points but also emphasizes more critical locations and vascular features within each time-point image, fusing them to obtain richer feature information. When tested on our constructed mCTA dataset, under the same training settings, the model we proposed achieved an accuracy rate that surpassed other classical image classification models by at least 5%, offering clear and practical recommendations to doctors, the barriers to diagnosing collateral circulation are significantly lowered, facilitating more accurate assessments.