Abstract
Digital chest x-ray images are the most widely used medical diagnostic tool Nowsaday. They helps in the diagnosis of diseases such as pulmonary tuberculosis and lung cancer. There are millions of chest radiographs produced by hospitals each year. But only less than 10% of the images were interpreted by radiologists. The rest of the images are often interpreted by general practitioners. But interpreting results by general practitioners is likely to be wrong. Several studies have shown that general practitioners may have chances to misinterpret the results including misdiagnose or missed diagnose up to 20-50% of the images with lung cancer lesions. In addition, the interpretation of results by a single radiologists still has a 5-9 percent chance of error. Artificial intelligence (AI) may be an option to assist in the screening of chest radiographs with lesions or abnormalities that must be reinterpreted by a radiologist. In this study, a database of chest radiographs from the tuberculosis and lung cancer database was created. For the development of AI systems using U-Net and SegNet neural networks, it was found that the SegNet model network yielded satisfactory screening results. Sematic segmentation using the SegNet model and using pre-tutorial VGG19, which added a dropout with a probability value of 0.4. After designing and teaching the network with various patterns until the level of accuracy is satisfactory. The AI was tested with testing images. For the test with 17 images, Small1I lesions was best identified by the AI with an average IoU accuracy of 90.23%, followed by cancer lesions with an average IoU of 89.45%. The lowest correctly classified lesion was cavity with an accuracy of only 5.71%. The AI was further tested with a set of 26 images with cancer lesions. The AI did not detect 4 malignant lesions. When tested with 10 images without lesions the AI incorrectly identified a couple of lesions in 2 images. From the results, this artificial intelligence should be implemented in the form of a Triage to screen images before consulting a radiologist. This can reduce the workload of radiologists and general physicians and thus may be of economic value in terms of radiologist and general practitioner wages and time.
