An aggregate method for thorax diseases classification

Abstract

A common problem found in real-word medical image classification is the inherent imbalance of the positive and negative patterns in the dataset where positive patterns are usually rare. Moreover, in the classification of multiple classes with neural network, a training pattern is treated as a positive pattern in one output node and negative in all the remaining output nodes. In this paper, the weights of a training pattern in the loss function are designed based not only on the number of the training patterns in the class but also on the different nodes where one of them treats this training pattern as positive and the others treat it as negative. We propose a combined approach of weights calculation algorithm for deep network training and the training optimization from the state-of-the-art deep network architecture for thorax diseases classification problem. Experimental results on the Chest X-Ray image dataset demonstrate that this new weighting scheme improves classification performances, also the training optimization from the EfficientNet improves the performance furthermore. We compare the aggregate method with several performances from the previous study of thorax diseases classifications to provide the fair comparisons against the proposed method.

A 178-clinical-center experiment of integrating AI solutions for lung pathology diagnosis

“…The comparison of AI to three radiologists on challenging-to-analyze cases from SIIM has demonstrated that AI can segment pneumothorax pockets more accurately than the radiologists, while the radiologists were more accurate in pneumothorax/no pneumothorax classification 66 . The results on the NIH database are around 0.80–0.98 AUC 17 , 19 , 20 , 23 . The superior results on the SIIM challenge, where the testing labels are not available to the algorithm developers, in contrast to the result on the NIH database suggest that binary pneumothorax diagnosis could be simpler than pneumothorax diagnosis as a part of a multi-disease analysis.…”

“…Due to its broad definition, infiltration is the most prevalent class in one of the most popular public lung X-ray databases – ChestX-ray8 from the National Institutes of Health (NIH) 16 . The availability of public data has summoned significant attention from the data science community to the automated infiltrate detection problem 6 , 17 – 34 . Most of the authors addressed the problem using classification neural networks usually with ResNet 16 , 29 , 31 , 33 , 34 and DenseNet 6 , 20 , 28 , 32 , 34 backbones.…”

“…Most of the authors addressed the problem using classification neural networks usually with ResNet 16 , 29 , 31 , 33 , 34 and DenseNet 6 , 20 , 28 , 32 , 34 backbones. The performance of the published solutions tested on the NIH database mostly lies in a narrow interval from 0.689 to 0.727 AUC 17 , 20 , 21 , 23 – 26 , 28 , 30 , 31 , 34 , 35 . Augmenting a classification network with lung field segmentation 28 or abnormality region enhancement 22 , 23 resulted in a 1–2% of accuracy improvement.…”

“…Pulmonary masses, defined as pulmonary opacities larger than 30 mm, are present in 2086 X-rays in the experiment database. Lung masses are included in the NIH database and their automated detection accuracy is currently estimated to be from 0.79 to 0.87 AUC 6 , 17 – 21 , 23 – 29 , 31 , 33 , 35 . Two studies used private X-ray datasets and were able to demonstrate the performance way above the results of other groups on the public NIH database with the AUC of 0.95–0.97 22 , 41 .…”

“…In all these papers, the UNet network was used for heart and lung segmentation. The standard end-to-end solutions, where networks are asked to predict the disease by directly analyzing the raw X-ray, resulted in AUC from 0.60 to 0.89 tested on the NIH database 17 , 18 , 20 , 23 – 29 , 31 , 33 , 34 . The results on alternative public databases such as CheXpert, MIMIC, OpenI, and PadChest, and their combination with private databases can reach 0.945 AUC 19 , 21 , 22 , 30 , 32 , 35 – 37 , 39 , 42 , 44 , 49 – 51 .…”