使用視網膜光學斷層掃描儀的醫學影像探討遷移學習在小資料集的醫學影像的效能

Efficacy of transfer learning on small datasets by using medical images from optical coherence tomography

陳家耀(CHIA-YAO CHEN)

指導教授 ： 蘇家玉

臺北醫學大學/醫學科技學院/醫學資訊研究所/碩士(2020年)

摘要

時至今日，視網膜光學斷層掃描儀的技術已被眼科醫師們廣泛使用於診斷眼科相關的疾病，例如脈絡膜的血管新生，糖尿病性黃斑部水腫和隱結堆積。然而，在實際情況下，設備的數量和經驗豐富的檢查人員通常是有限的。因此，電腦輔助診斷系統的使用，可能有助於這些醫學影像相關的任務。但是，在現實世界裡，能作為醫學研究的樣本數量可能也相對有限。而現今，解決小樣本問題的一個潛在方法，是利用遷移學習。因此，本研究旨在通過使用來自於龐大的視網膜光學斷層掃描儀的影像資料集中，取出的較小數量的子集，來評估遷移學習的功效。這些實驗表明，通過使用基於遷移學習構建的模型，數百張影像可能足以在疾病和健康分類上提供良好的性能。例如，在256個異常圖像組別，有一半的實驗結果其預測準確率超過90%。而在脈絡膜的血管新生，糖尿病性黃斑部水腫和隱結堆積中，這些預測準確率的平均值分別為94.55％，92.44％和90.77％。即使是從僅數個但具有多個訓練週期的樣本建立的模型，也有可能像是從數千個影像中獲得的模型一樣具有較高的功效。例如，在8個異常圖像組別，經過1024個訓練週期後，脈絡膜的血管新生的平均預測準確率為94.34％，糖尿病性黃斑部水腫為79.13％，隱結堆積則為78.83％。儘管在本研究中，電腦計算所耗費的時間並不是主要的研究目標，但是從預先訓練的神經網絡中進行的遷移學習，也能用較省時的方式構建有效的模型。從這項研究中獲得的見解，可能有助於樣本數量受限的醫學影像的相關研究。這也是機器學習未來在醫學領域中的應用與研究的重要客題。

關鍵字

視網膜光學斷層掃描儀 ； 遷移學習 ； 小資料集 ； 醫學影像

並列摘要

Nowadays, optical coherence tomography is used widely by ophthalmologists for diagnosing diseases, such as choroidal neovascularization (CNV), diabetic macular edema (DME), and drusen (DRU). Nevertheless, availability of equipment and experienced imaging personnel is often limited. Therefore, the utility of computer-aided diagnosis systems may be helpful in those medical image interpretation tasks. However, among the real world, the sample size for medical research can be relatively restricted. One potential way to solve the small sample size problems is the utilization of transfer learning. Thus, this study intends to assess the efficacy of transfer learning by using smaller subsets from a larger optical coherence tomography image dataset. These experiments revealed that hundreds of images could be adequate to provide good performance on classification of disease and normal by using the models built from transfer learning. For example, with 256 abnormal images, half of the results about accuracy could exceed 90%. Those results of accuracy were with the average of 94.55%, 92.44%, and 90.77% in CNV, DME and DRU, respectively. Even models established from as few as dozens of samples with multiple training epochs could have a chance to achieve high efficacy as those from thousands of images. For instance, in the group of eight abnormal images after 1024 epochs of training, the average accuracy is 94.34% in CNV, 79.13% in DME and 78.83% in DRU. And transfer learning certainly allows us to build efficient models in a timesaving way, although computation time is not a leading topic in this study. The insights gained from this study may be of assistance to the investigations about medical images but with limited sample size. This is also an important issue for future research about the application of machine learning in medical field.

並列關鍵字

optical coherence tomography ； transfer learning ； small datasets ； medical images

參考文獻

References

[1] Haring CM, Cools BM, van der Meer JW, Postma CT. Student performance of the general physical examination in internal medicine: an observational study. BMC Med Educ 2014;14:73. https://doi.org/10.1186/1472-6920-14-73.

[2] Méndez A. Optics in Medicine. In: Al-Amri MD, El-Gomati M, Zubairy MS, editors. Opt. Our Time, Cham: Springer International Publishing; 2016, p. 299–333. https://doi.org/10.1007/978-3-319-31903-2_13.

[3] Helmholtz H von. Handbuch der physiologischen Optik. Hamburg: Voss; 1896.

[4] Dimmer F (Friedrich). Die Photographie des Augenhintergrundes. Wiesbaden : Verlag von J.F. Bergmann; 1907.