全玻片高通量影像顯微術以實現精準醫學：腫瘤比例評分之自動化演算

Whole Slide Image and Deep Learning Based Technique for Determination of Tumor Proportion Scores

廖信宇(Xin-Yu Liao)

指導教授 ： 董成淵

國立臺灣大學/理學院/物理學研究所/碩士(2020年)

https://doi.org/10.6342/NTU202002888

摘要

PD-L1表現率之腫瘤比例評分為現今評定免疫檢查點阻斷療法於癌症病患個案是否適宜之重要生物指標。以臺灣健保系統對免疫療法之給付規定為例，頭頸部鱗狀細胞癌及非小細胞肺癌患者需於該評分被評定高於50%，始能通過給付條件；然而現今腫瘤比例評分係由病理科醫師進行人工計數判定，以口腔癌為例，一個大於1公分的樣品平均包含100,000個以上細胞，故現行流程將十分費時，且人力計數方式將不甚精準。為改善該評分流程，以期符合精準醫學之方針，本研究中，我們以高解析度白光顯微鏡搭配精密樣品載物平台，並將影像訊號透過工業彩色相機擷取，藉LabVIEW編譯之系統控制軟體以實現自動化掃圖，並將子影像透過ImageJ輔助拼接出可供分析之全玻片影像。由全玻片影像提供之高通量資訊，我們透過深度學習演算法自動計算出患者之腫瘤比例評分，與臺大醫院病理科醫師估計值相差低於10%，可望於未來發展為數位病理輔助儀器以實現精準醫學；另外，本研究發現在同一塊腫瘤切片中，相隔僅3毫米之兩片切面薄片之評分竟可以從34% 驟增至62%，足以證明腫瘤比例評分亟需數位病理方法輔助以獲得更大數據之統計結論，而非繼續採現行之人工計數方式。由於全玻片影像之品質將大大影響腫瘤比例評分正確性，故本研究亦探討實驗中光學系統及亮度平坦度之校正及染劑分子濃度對彩色影像畫素組成之影響。最後，於實驗中所取得之全玻片影像資料將可供作為人工智慧判讀病人療效之訓練資料，展望更完整且精準的數位病理判讀研究。

關鍵字

全玻片影像 ； 免疫檢查點阻斷療法 ； 腫瘤比例評分 ； 數位病理學 ； 精準醫學

並列摘要

Currently, Tumor Proportion Scores (TPS) of PD-L1 expression is the clinical bio-marker for immune checkpoint inhibitor therapy. For the health insurance system in Taiwan, only cases with TPS≧50% for head and neck cancer or NSCLC are covered; however, TPS is clinically determined by pathologists using manual counting to estimate the percentage of viable tumor cells with PD-L1 labeled membrane. In the case of a 10-mm sample, more than 100,000 cells can be present and it is time-consuming and lacks precision and accuracy from manual counting. To improve the scoring process, in this research, an auto-scanning instrument composed of a bright field microscope (TE2000-U), a high-resolution prior stage, and an CCD-camera was constructed through the use of LabVIEW based software, which enables whole-slide imaging capability of biopsy slices through ImageJ stitching function. A deep learning based algorithm was constructed to process high throughput WSI data from which we can automatically calculate patients’ TPSs. Our results are within 10% error to the scores evaluated by a pathologist in NTUH. Furthermore, we also find that the TPS of two adjacent tissue sections separated 3 mm apart can drastically increase from 34% to 62%. Therefore, current procedure for assessing TPS needs digital pathology analysis of large data sets to achieve more significant statistical results. Besides, the accuracy of TPS is greatly affected by the image quality; therefore, we would also probe into the calibration of optics systems and the flatness of intensity. Ultimately, the image data acquired in this research would be provided as the training data for building a more complete AI software to estimate the curative effect more precisely in the future.

並列關鍵字

whole slide imaging ； immune checkpoint inhibitor therapy ； Tumor Proportion Score ； digital pathology ； precision medicine

參考文獻

CECIL H. FOX. et al., “Formaldehyde Fixation”, The Journal of Histochemistry and Cytochemistry. 5R0344A (1985).

Musumeci, Giuseppe. "Past, present and future: overview on histology and histopathology." (2014): 5.

G. Rolls, “An Introduction to Specimen Processing.” Leica Biosystems

Bancroft, John D., and Marilyn Gamble, eds. Theory and practice of histological techniques. Elsevier Health Sciences, 2008.

Wang J et al. (2017) Journal of Hematology Oncology 10(1):34 DOI: 10.1186/s13045-017-0403-5