【恭賀】數據所許志仲老師 指導碩生-陳劭寧、戴勝捷獲得全球第四佳績!!

恭喜 成大數據科學研究所許志仲老師 指導碩生 陳劭寧、戴勝捷 參加The International Association for Pattern Recognition (IAPR) 旗下國際會議ICIAP舉辦Covid-19-Infection-Percentage-Estimation-Challenge,從176參賽隊伍中,獲得全球第四的佳績!

本次競賽是從給定的 Computed Tomography (CT) Scan 中,評判每一個 Slice 的Coivd-19 相關病徵的分數,其中每個 Slice 的感染分數是由兩位專科醫生所標註取平均的,其中訓練資料僅有 132 CT,相對來講相當少,再加上我們僅有每個Slice的影像,也無CT相關的資訊 (如切片厚度等),所以要利用深度學習模型學習出來相對不易。兩位同學雖然經驗不多,但最終仍努力提出利用Swin-Transformer,再加上特徵加強、多任務學習、以及Augmented ROI,學習出一個強健不易受影響的模型。雖然初賽階段僅有第六名,但因為我們發展的模型相對更為穩定,因此最終在決賽時,面對新一批的資料仍舊維持同樣的效能,因而反轉獲得全球第四的佳績,實屬不易!

Congratulations to Shao-Ning Chen and Shen-Chieh Jerry Dai, my graduate students from the Institute of Data Science in NCKU, who participated in the Covid-19-Infection-Percentage-Estimation-Challenge held by ICIAP, an international conference under The International Association for Pattern Recognition (IAPR), and won the fourth place in the world!

This challenge aims to judge the infection scores of Coivd-19-related symptoms of each slice from a given Computed Tomography (CT) Scan. The number of the training data is only 132 CT, which is relatively small. We propose a Swin-Transformer-based approach, coupled with feature enhancement, multi-task learning, and Augmented ROI, to learn a robust model. Related information could be found at