论文部分内容阅读
Leukemia is a type of blood cancer that affects both children and adults,and it is one of the leading causes of death all over the world.Acute lymphoblastic leukemia（ALL）is the most wellknown types of leukemia,which develops in the human bone marrow due to producing more immature white blood cells（WBCs）.The techniques like bone marrow examination and manual blood checks that have been used a long time ago are often slow and error-prone sometimes do not give the desired results that we need.Since,with the coming of Convolutional neural networks（CNNs）and the promising results demonstrated in several applications,deep learning techniques for ALL classification are becoming more popular.A computer-aided cancer diagnosis（CAD）system has gained significant attention with the recent advancement in CNNs,because they are cost-effective and easily deployable.In this paper,we proposed two CNN based models to classify ALL cancer.Our first model is based on using an attention mechanism with a deep neural network to address the classification of ALL cancer.We used the normal vs.malignant cells（C-NMC）2019 challenge dataset.A few researchers have used C-NMC 2019 datasets to classify ALL cancer into healthy and cancer classes,but they did not consider the subject level variability.We divide the dataset into 7-Folds based on subject-level variability.This distribution splits the dataset into different folds and make sure that no two folds have the common images of the same subject.Furthermore,to overwhelm the problem of complex morphological similarity between cancerous and healthy cell images,we deployed an attention mechanism,namely Efficient Channel Attention（ECA-Net）in-plane visual geometry group from oxford（VGG16）,to extract deep features from the images and give better classification results.Since CNN models required a lot of data for training,different augmentation techniques are applied to C-NMC 2019 challenge dataset.In order to show the effectiveness of our proposed model we also demonstrated different feature maps generated from the first and second blocks of VGG16 with and without the attention module.The demonstrated feature maps clearly concluded that the ECA module forces the module to focus on the cell area,shape,texture and identify high-level features.Our Attention-based CNN provides sufficient results for the classification of ALL into two classes,but there are some problems associated with our model.The first problem is we only utilized C-NMC 2019 challenge dataset for the classification,and we did not check our model on other datasets such as ALL-Image Database（IDB-1）and ALL-Image Database（IDB-2）.The second problem associated with our proposed model is that it provides comparatively high falsepositive rates in the classification phase.Finally,there is also a need for a model based on the recently proposed deep neural network to improve accuracy.Therefore,we also proposed a weighted ensemble model based on the information bottleneck for attribution（IBA）theory to classify ALL cancer into healthy and cancer cell images.We used recently proposed pre-trained CNN models such as Efficient Net,Dense Net121,and Res Net50.In order to check the effectiveness of our proposed model,we used three publicly available datasets.First,we train and test plain Efficent Net,Dense Net,and Res Net50 models to classify ALL diseases.Secondly,we used IBA theory with our proposed models,which helps the model in extracting and refining feature representation relevant to classification,and also aids in the final accuracy of CNN.Finally,since deep learning models differ in design and complexity,they may not provide the same results for a specific problem domain.Some models will perform better some will give poor results.As a result,it would be beneficial if we assigned higher weights to the models that are performing better,allowing us to utilize the efficiency of a model which is contributing more to the final accuracy as compared to the model that performs poorly.The obtained findings showed that the proposed model could be used to identify ALL and might also assist pathologists in diagnosing ALL cancer in its early stages,potentially saving more lives.