應用輕量化深度學習模型進行非小細胞肺癌病理影像分類

Taipei Medical University Institutional Repository > 醫學院 > 人工智慧醫療碩士在職專班 > 博碩士論文 > Item 987654321/65006

請使用永久網址來引用或連結此文件: http://libir.tmu.edu.tw/handle/987654321/65006

題名:	應用輕量化深度學習模型進行非小細胞肺癌病理影像分類 Utilizing Lightweight Deep Learning Models for Non-Small Cell Lung Cancer Pathological Image Classification
作者:	簡菘宏 Chien, Sung-Hung
貢獻者:	醫學院人工智慧醫療碩士在職專班林于翔
關鍵詞:	輕量化深度學習模型;非小細胞肺癌;病理影像分類;邊緣運算 Lightweight deep learning model;Non-small cell lung cancer;Pathological image classification;Edge computing
日期:	2024-07-16
上傳時間:	2025-01-06 09:19:29 (UTC+8)
摘要:	非小細胞肺癌（NSCLC）為肺癌中最常見的類型，早期診斷和分類對於改善患者預後相當重要，然而，傳統的病理診斷方法不僅耗時，還容易受到人為因素的影響，因而導致診斷結果有較大的變異性。近年來，有許多深度學習技術導入到病理影像分類領域，並獲得了巨大的成功，因此，利用深度學習輔助NSCLC的診斷成為了一個重要的研究議題。有鑑於此，本研究提出了一種新型輕量級深度學習模型以輔助NSCLC的診斷，除此之外，本研究還進一步探討其在邊緣運算設備的臨床應用潛力。本研究提出了兩種名為MobileLungNet和MobileLungNe-Lite的新型輕量級深度學習模型，並使用包含15,000張肺部病理圖像的LC25000數據集進行了模型訓練和性能評估。實驗結果顯示，本研究提出的MobileLungNet在FP32和FP16壓縮格式之下，均達到了99%以上的高分類準確率。除此之外，在進行INT8量化後，MobileLungNet之分類準確率仍然優於其他同類型的經典輕量級深度學習模型，如MobileNet-v1和MobiHisNet。在實際的邊緣設備部署測試中，我們選擇了廣泛使用的Raspberry Pi作為邊緣運算測試平台，實驗結果顯示，本研究所提出之MobileLungNet-Lite輕量級模型在INT8壓縮格式下，僅需200毫秒以內就能完成一次診斷，顯示了本研究所提出的新型輕量級深度學習模型，具有未來於臨床端進行邊緣運算應用的潛力。 Non-small cell lung cancer (NSCLC), as the most common type of lung cancer, requires early diagnosis and classification to improve patient prognosis. However, traditional pathological diagnostic methods are not only time-consuming but also susceptible to human factors, leading to significant variability in diagnostic results. In recent years, many deep learning techniques have been introduced to the field of pathological image classification with great success, making the use of deep learning to assist NSCLC diagnosis an important research topic. In light of this, this study proposes a novel lightweight deep learning model to aid in NSCLC diagnosis and further explores its potential for clinical applications on edge computing devices.This study proposes two novel lightweight deep learning models named MobileLungNet and MobileLungNet-Lite, which were trained and evaluated using the LC25000 dataset containing 15,000 lung pathology images. Experimental results show that the proposed MobileLungNet achieved a high classification accuracy of over 99% in both FP32 and FP16 compression formats. Moreover, after INT8 quantization, MobileLungNet's classification accuracy still outperformed other classic lightweight deep learning models of the same type, such as MobileNet-v1 and MobiHisNet.In practical edge device deployment tests, we chose the widely used Raspberry Pi as the edge computing test platform. Experimental results show that the lightweight MobileLungNet-Lite model proposed in this study can complete a diagnosis within 200 milliseconds under the INT8 compression format, demonstrating the potential of the novel lightweight deep learning model for future clinical edge computing applications.
描述:	碩士指導教授：林于翔口試委員：林于翔口試委員：彭徐鈞口試委員：蘇家玉
附註:	論文公開日期：2029-07-21
資料類型:	thesis
顯示於類別:	[人工智慧醫療碩士在職專班] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	0	檢視/開啟

在TMUIR中所有的資料項目都受到原著作權保護.

TAIR相關文章

資料載入中.....