Comparison of segmentation methods for the detection of breast cancer using thermal images

Breast cancer is the most diagnosed cancer among woman and one of the deadliest. Early detection is key in bettering the chance of a recovery, it ranges from a survival change of 100% at stage 1 to 22% at stage 4. There are several methods designed for the early detection of breast cancer, with the gold standard being mammography. There is a need for the development of non-invasive and cost-effective methods. This has led to other methods being investigated. One such method is thermography, which analyses heat patterns captured using an infrared camera. It then looks for asymmetry between the left and right breasts as an indication of abnormality. Segmenting out the breasts correctly is a vital part of the detection process. There are many ways to extract the breasts. This study compares five segmentation methods, namely: distance-based cropping (Method 1), manual zoom cropping (Method 2), semi-automatic segmentation (Method 3), fully manual segmentation (Method 4), and fully automatic segmentation (Method 5). Once the breasts are extracted via each method statistical and textural features are calculated and reduced to only the significant ones. From this a simple feed forward neural network is trained using leave out one cross validation for robustness. Method 1 has 63.33% accuracy, Method 2 had 76.67%%, Method 3 had 88.33%, Method 4 had 90%, and Method 5 had 80%. Method 5 had 6 unsuccessful segmentations due to it being unable to detect the inframammary folds, due to this semi-automatic segmentation (Method 3) is the recommended approach due to its high accuracy and a degree of automation which saves time.