Research On Fine-grained Image Classification Method For Medical Image Computer Aided Diagnosis

Compared with other methods,medical imaging is more intuitive and easier to obtain,and it is widely used in the diagnostic process of various diseases.Computer-aided diagnosis based on medical images can save the energy of doctors and make early screening easier,which is of great significance for the prognosis of patients.Medical images based diagnosis often requires classification of images at sub-class level,which is essentially a fine-grained image classification problem in natural images analysis.Motivated by the fact,we present a finegrained image classification method to boost the classification performance in the context of localization.Our proposed method works in a weakly-supervised fashion,which only takes as input image-level class labels,without the necessity of expensive part annotations.Due to the limited amount of samples in medical data,we improve the classification performance by extracting and fusing global and local information based on the two-level attention mechanism.An image-level convolution neural network(CNN)and a patch-level CNN are involved to deal with the images and patches.The final classification result is the integration of output from CNNs in both image-level and patch-level.The analysis of medical images requires expertise and experience,which makes the precise part annotation expensive.Instead of manual annotations,we locate and crop the discriminative patches based on the bottom-up perspective.A series of candidate bounding boxes are proposed first and filter with the lesion information extracted by the neural network visualization technology.The remaining bounding boxes are considered as discriminative patches which can boost the performance.Experiments on real clinical data demonstrate that the proposed method can achieve promising performance.In the context of otitis media diagnosis with endoscopic tympanic membrane images,we observed a stable improvement from the integration of lesion information.Motivated by this,we construct the LSID dataset to further study on laryngoscopic image analysis for computer-aided diagnosis of laryngeal cancer.The effects of different fusion strategies,different lesion information generation methods,different region proposal methods and the number of patches on the classification accuracy were explored.We also compared it with a series of existing methods.Experiments show that the proposed method performs well on different types of medical image classification problems and achieve a good generalization ability.