基於多示例學習之高效階層式感興趣區域檢測於全切片影像中的嗜酸性白血球數量估計

Abstract

嗜酸性慢性鼻竇炎（Eosinophilic Chronic Rhinosinusitis, ECRS）是一種涉及鼻部的慢性免疫系統疾病，其特徵是免疫系統失調導致異常數量的嗜酸性白血球聚集在鼻部和鼻竇的部分區域。可能的症狀包括鼻塞、鼻瘜肉的生長、嗅覺衰退以及鼻涕倒流等等。診斷ECRS需要去計算鼻部組織切片的嗜酸性白血球數量。透過計算數量，除了可以得知確診與否，也可以以白血球數量數值的高低來作為目前的恢復狀況。

目前基於全切片醫學影像之研究主要集中於腫瘤分類與細胞圖像分割。近年來透過深度學習與多示例學習的發展得到了重大的改進。然而在嗜酸性白血球數量估計的題目中，目前鮮有研究進行。其最主要的原因在於全切片醫學影像資料的不足以及嗜酸性白血球圖片標記的不易。因此我們基於多示例學習在腫瘤分類中的重大突破，提出了一個快速選取感興趣區域的方法來加速。此外，為了增進醫師在做嗜酸性白血球圖片標記的效率，本研究提出了以補丁形式作為細胞標註單位的方式來減少醫師做資料標記的負擔。本研究使用台大醫院耳鼻喉科所提供之資料集來作為訓練與測試，最終達到在全切片影像中快速得到分類結果與高準確率的研究成果。

Eosinophilic Chronic Rhinosinusitis (ECRS) is a chronic immune system disorder affecting the nasal passages. Its hallmark is an immune system imbalance leading to an abnormal accumulation of eosinophils in specific areas of the nose and sinuses. Symptoms may include nasal congestion, the growth of nasal polyps, diminished sense of smell, and postnasal drip. Diagnosing ECRS involves calculating the quantity of eosinophils in nasal tissue slices. This quantitative assessment not only aids in confirming the diagnosis but also serves as an indicator of the current recovery status based on eosinophil count levels.

Current research in whole medical imaging primarily focuses on tumor classification and cell image segmentation. Advances in deep learning and multi-instance learning in recent years have significantly improved these areas. However, there is a scarcity of research in the estimation of eosinophil counts, primarily due to insufficient data in whole-slide medical images and the difficulty in annotating eosinophil images. Building on the breakthroughs in multi-instance learning for tumor classification, our study proposes a rapid region of interest selection method to expedite the process.

Additionally, to enhance efficiency in annotating eosinophil images, this research suggests using patch-based cell labeling to reduce the burden on medical professionals. The dataset provided by the Department of Otolaryngology at National Taiwan University Hospital is utilized for training and testing purposes. Ultimately, our study aims to achieve rapid classification results with high accuracy in whole-slide images of eosinophilic chronic rhinosinusitis.