Semantic segmentation of medical images with deep learning models is rapidly being developed. In this study, we benchmarked state-of-the-art deep learning segmentation algorithms on our clinical stereotactic radiosurgery dataset. The dataset consists of labeled brain CT from 1000 patients with various brain lesions (pituitary tumors, meningioma, schwannoma, brain metastases, and arteriovenous malformation). This study demonstrates the strengths and weaknesses of deep-learning algorithms in a fairly practical scenario. We compared the model performances concerning their sampling method, model architecture, and the choice of loss functions, identifying suitable settings for their applications and shedding light on the possible improvements. Evidence from this study led us to conclude that deep learning could be promising in assisting the segmentation of brain lesions even if the training dataset was of high heterogeneity in lesion types and sizes.

使用胃鏡檢查之影像，以及胃黏膜組織學確診資料，運用人工智慧技術分析這些影像，以期瞭解胃相、預測組織學上的變化，希望此分析系統能量化民眾未來胃癌的風險，以協助臨床醫師判斷哪些病患需要胃鏡的定期追蹤檢查，建立起的風險預測模型，不僅能在大型醫院中運用，更期待能在專業人員不足的偏鄉場域（如馬祖）能進行驗證及落地使用，以促進整體醫療水平的進步。

Hematopoietic stem cell transplantation (HSCT) is currently widely used as curative treatment for acute leukemia, lymphoma, and other hematological diseases, such as aplastic anemia or thalassemia major. Allogeneic HSCT has the curative ability for its immunotherapy characteristics from the donor cells against recipient’s malignancy, so call the graft-versus-leukemia/lymphoma (GVL effect). However, the usual concurrent acute or chronic graft-versus-host disease (aGVHD and cGVHD) remains one of the major causes of morbidity and mortality after allogeneic HSCT. Previous studies usually analyze severity in individual organs and overall disease severity of GVHD, but did not combine different organs for phenotypic clinical subgrouping. Machine learning has the advantage to find specific patterns in clinical data that may be missed by the traditional approach and observation. The aim of this study is to use machine learning to cluster clinical phenotypes of organ involvement in acute GVHD and investigate its prognostic role in patient received allogeneic HSCT in Taiwan.

ICD-10(The International Statistical Classification of Diseases and Related Health Problems 10th Revision)診斷代碼約69,000個、處置代碼約72,000個，目前需仰賴疾病分類專業人員閱讀病歷資料進行分類，本計畫擬開發智慧編碼系統輔助編碼。運用NLP技術分析病歷內容，並導入專家知識，產生ICD-10疾病分類編碼建議，輔助編碼作業，進而輔助醫師撰寫病歷提升病歷品質。

使用彩色眼底攝影這樣一個快速且相對便宜的方式，幫助我們在白內障手術前評估病患的視網膜，幫助評估白內障遮擋、過於細微或還未被證實影響視力潛能的病灶。此外，眼軸過長容易造成早發性以及程度較嚴重之白內障，在白內障術前檢查時眼軸長亦是個重要指標，因此本研究在利用彩色眼底攝影的同時也建立機器學習模型以預測眼軸長，以期未來在白內障手術的評估能達到更精準的醫療。

偏鄉醫療資源不足，為人工智慧落地運用的最佳場域，本計劃根據十年以上的馬祖與台大內視鏡影像，以及慢性胃炎病理診斷，發展出根據影像預測病理診斷的人工智慧模型，準確率達到9成，可以預測痿縮性胃炎及腸上皮化生，在偏鄉進行驗證。

國際疾病分類第十版ICD-10(International Classification of Diseases, 10th Revision)診斷代碼約69,000個、處置代碼約72,000個，需仰賴疾病分類專業人員閱讀病歷資料進行分類，本案開發智慧編碼輔助系統。運用自然語言處理技術，分析病歷紀錄內容(如出院診斷、手術紀錄…等)，並導入疾病分類專家知識，對於不同科別、不同疾病，以及特定的病歷撰寫方式進行細部優化，讓機器學習模型進一步提升效能，為每次住院個案演算出ICD-10-CM診斷碼及ICD-10-PCS處置碼之編碼建議，輔助編碼作業，進而輔助醫師撰寫病歷提升病歷品質，並建立AI文字資料處理及建模機制。

使用2016~2018年之流感病患電子病歷及X-Ray，將資料分為測試資料及以及驗證資料集進行模型訓練；本研究總共建立兩個模型，其一是DeepFlu_XR，以X-ray資料進行類神經網路建模，為胸腔X-ray嚴重度評分，其二是DeepFlu_fuison model，以結構化資料及DeepFlu_XR score進行機器學習，並預測死亡機率，為臨床做一個流感嚴重度分級系統。

本計畫使用數位病理影像，分析乳房切片診斷非典型乳腺增生及乳管原位癌之病患之病灶風險。影像模型可輸出切片組織總量、病灶含量、核分級、壞死程度、以及基質反應性程度。人工智慧判讀結果與病理科醫師判讀一致性高。結合臨床及影像資訊，模型對於預測後續廣泛性切除檢體是否有侵犯癌，可達近八成的準確度，並經獨立年份分割資料集測試。人工智慧模型可優化乳房癌前病變之風險評估。未來結合臨床試驗成果，可望為乳房癌前病變之患者提供治療導引。