醫學眼底影像中不平衡多標籤分類的準確性提升

Abstract

近年來，深度學習技術在醫學影像研究領域中的應用不斷深入，目的是提升 對各種病症的診斷準確度。當前面臨的主要挑戰是，醫學影像分類任務大多僅聚 焦於單一標籤，缺少對影像的全面診斷能力，尤其是在處理數據不平衡的多標籤 分類問題時。這些挑戰涉及到不同病變類型之間的資料量懸殊以及潛在的病變相 關性，這些因素都可能影響模型性能。 為應對這些問題，本篇論文開發了一款多標籤深度學習模型，對當前先進的 影像識別模型進行改進，使其適用於多標籤醫學影像分類任務。在資料前處理階 段，本研究設計了一套針對類別不平衡問題的數據增強演算法，利用公開醫學眼 底影像進行多標籤分類。本研究著重於解決眼底病變類型數量不平衡的問題，以 提升模型的準確性和泛化能力，從而在自動識別和分類眼底疾病上達到最佳性 能，實現對多種病變的準確診斷。 最後，通過公開醫學眼底資料集上進行訓練與效能評估，與現有的多標籤模 型進行比較，實驗結果顯示，我們的模型及其數據增強方法顯著提高了準確率， 且性能超越了現有的先進模型。

In recent years, the application of deep learning techniques in the field of medical imaging research has continued to deepen, with the goal of enhancing the diagnostic ac­ curacy for various diseases. A major challenge currently faced is that most medical image classification tasks focus only on single labels, lacking comprehensive diagnostic capabili­ ties for images, especially when dealing with imbalanced data in multi­label classification problems. These challenges involve disparities in data volumes among different lesion types and potential interrelations among lesions, which can impact model performance. To address these issues, this paper develops a multi-­label deep learning model that im­proves upon current advanced image recognition models, making them suitable for multi­ label medical image classification tasks. In the data preprocessing stage, this study designs a set of data augmentation algorithms specifically for class imbalance issues, using public medical fundus images for multi­label classification. This research focuses on resolving the issue of imbalanced numbers of fundus lesion types to enhance the accuracy and gen­eralizability of the model, thereby achieving optimal performance in the automatic identi­ fication and classification of fundus diseases and realizing accurate diagnosis of multiple lesions. Finally, by training and conducting performance evaluations on the public medical fundus datasets, and comparing with existing multi-­label models, the experimental results demonstrate that our model and its data augmentation methods significantly improve ac­ curacy and outperform existing advanced models.