This paper describes our participation at the ImageCLEF 2015 Medical clustering task. The task is about clustering digital x-ray images into four groups with regard to the body parts. We experimented with deep convolutional neural network (GoogLeNet), finetuning pretrained models for ImageNet dataset. Experimental results showed competitive performance with other top-performing runs.
In this paper, we describe our participation at the ImageCLEF 2015 [1, 2] medical clustering [3] task. Given digital x-ray images of various body parts, task purpose is clustering images into four different body parts: head-neck, upper-limb, body and lowerlimb. For a detailed introduction of the task, please see the overview paper of this task [4]. In this study, we experimented with deep convolutional neural network (CNN). In recent years, CNN showed quite effective performance in image classification tasks [5]. We formulated this task as an image classification among four different body part labels. We experimented with GoogLeNet which was used in recent ImageNet Challenge [6]. GoogLeNet incorporates Inception module with the intention of increasing network depth with computational efficiency.
We randomly divided training set into five-fold. Images from one fold is used as validation set, and images from other four fold is used as training set. We finetuned GoogLeNet pretrained on ImageNet dataset (initial learning rate 0.001; batch_size:40). 90 degree rotation (90’, 180’, 270’ and 360’) of images, mirroring (random left-right flipping of image) and image cropping (random cropping 224 x 224 image window out of 300 x 300 resized image) is applied for input data augmentation. Our trained CNN models scored 0.89~0.93 top-1 accuracy in our validation set.
We trained five separate CNNs. Five ranked list is combined into single ranking using Borda-fuse method [7]. Only top-ranked body parts are marked as output in test set. Borda-fuse method combines individual ranks without utilizing score. Combining multiple CNN classification output is considered to be effective to cope with CNN’s variance. We postponed experimenting with other metasearch techniques such as CombSUM [8] to the future work. 3
In GoogLeNet, there are three output layers (loss1, loss2 and loss3), two of them (loss1 and loss2) is located in the middle of layer hierarchy. We used these three layers per each run. Our run SNUMedifo1 corresponds to the lowest output layer (loss1). SNUMedinfo3 corresponds to the uppermost output layer (loss3).