Computer Aided Diagnosis Via Multi-scale Deep Learning Algorithm And Few-shot Learning Method

Currently,machine learning has been recognized as an efficient tool in fields of medicine,like medical images processing.Deep learning allows automatic and fast feature extracted from raw data,instead of traditional expert knowledge-based feature extraction.Some of these machine learning based methods achieve excellent performance.However,these methods are heavily dependent on large labeled datasets for supervised training.In most instances,it is not feasible to obtain large-scale labeled datasets for specific medical conditions.Previous studies using machine learning algorithms have commonly applied data augmentation methods.While this improves performance on the initial dataset,it usually fails to capture the underlying complex variations.A strong generalization ability is indispensable in medical image classification tasks.Another widely used technique is the few-shot learning method,which seeks good generalization on problems with a very limited labeled dataset,typically containing just a few training samples of the target classes.In medical image research,it is often not enough to extract a single feature.Multi-scale learning methods can effectively extract more feature information.This paper combines few-shoting learning method and multi-scale learning to construct a new type of deep network model to study the problem of medical image classification.The main content of the paper is as follows:(1)A novel framework is proposed for SNHL classification task.In the framework,we combined CNN and the comparison classifier strategy.First,CNN was utilized as a deep feature extractor.The advantage of this deep feature extractor is that it automatically explores the underlying hierarchal relationships within the data.This character-istic can effectively replace the laborious process of hand-crafting features.The few-shot learning algorithm was then used to replace data augmentation as a method to resist the lack of MRI data and increase generalizability.Finally,target SNHL MR images are classified into three categories:left-sided SNHL(LSHL),right-sided SNHL(RSHL),and healthy controls(HC).This novel framework can accurately classify our SNHL dataset after pre-training,which significantly improves efficiency and demonstrates excellent generalization ability.(2)Use multi-scale deep learning methods for computer-aided diagnosis.As a representative algorithm of deep learning,convolutional neural network has strong feature learning ability.This algorithm can effectively extract high-frequency and low-frequency features of images,and is very suitable for medical images that require pixel-level analysis.This paper classifies the pathological images of cervical neuroendocrine cancer based on multi-scale deep learning algorithms.The experimental results show that the multi-scale deep learning algorithm can extract image features well for image classification tasks.