Abstract:
Finding tumor areas on breast ultrasonography pictures has long been a fascinating subject. The
complex architecture of breast and the presence of noise in ultrasound images sometimes make it
impossible for traditional handcrafted feature-based approaches to produce adequate results. The
accuracy of finding objects has significantly improved with the recent developments in deep
learning, particularly for generic detection of objects. Moreover, most models currently lack
efficient optimization for the algorithm's structure, which incurs high processing costs during
training and deployment. This paper offers a variety of image processing techniques for breast
tumor classification. Finding breast tumor is the study's main objective. In order to accurately
detect and classify images as benign or tumor, a number of methods and algorithms have been
developed. In this particular experiment, ultrasound pictures of three different class kinds—
malignant, benign, and normal—were combined with dl-based models. The prediction and
detection of tumor pictures is done using five models: InceptionResNetV2, InceptionV3, VGG16,
VGG19, and DenseNet169, to classify breast tumor stages. Finally, the results of the approach are
assessed using two different measures of efficiency. Four possible outcomes—TP, TN, FP, and
FN—are used in the first reliability set, a performance evaluation for the stages of tumor that
follow. We next apply the above algorithms to analyze the accuracy of each type of breast tumor
in mistake scenarios. The InceptionResNetV2 approach, which I suggest, allows for the
autonomous recognition of breast tumors with an accuracy rate of 91.82%.
