soft classification models combined with hyperspectral imaging for diagnosis of breast cancer

Abstract: breast cancer is one of the known cancers. among every 8 females, one will suffer from malignant breast tissue growth, during their lifespan [1]. diagnosis of breast cancer in the early stages is an important matter and can lead to full recovery. therefore, finding a solution for diagnosis in the early stages, with a high sensitivity is a great concern. currently biopsy is the standard method for breast cancer sampling [2]. there are some drawbacks for this method, for instance the breast tissue has to be fully removed in order to determine the margin of the tumor and as a result some parts of the healthy tissue will be in the biopsy sample. hyperspectral imaging (hsi) is a novel method in medical field and it has shown promising results in the diagnosis of cancer (benign and malignant) [3]. hsi offers information about spatial and spectral properties of the sample that determines the distribution map and the identity of the present components in the sample [4]. in the present contribution, vis-nir hsi combined with different hard and soft classification methods is proposed for diagnosis of breast cancer. in this regard, 56 breast cancer biopsy samples from healthy and cancerous tissues were provided. it should be pointed out that the cancerous biopsy samples were from grade 2 and grade 3. vis-nir hsi images of the samples were acquired in the spectral range of 400-950 nm. to explore the similarities and dissimilarities among samples, initially, principal component analysis (pca) was used which showed a distinction between healthy and cancerous samples. the data was then analyzed with soft models of partial least squares-discriminant analysis (soft pls-da) and soft independent modelling of class analogy (simca). the goal was to explore the ability of these methods in comparison with hard classification methods to model different classes. the classification figures of merit in terms of sensitivity and specificity were promising for both algorithms. as an instance, the sensitivity values for healthy, grade 2 and grade 3 samples were respectively 89.5%, 84.4%, 92.5%. additionally, the specificity values were 87.0%, 72.1% and 89.2% for healthy, grade 2 and grade 3 samples, respectively. comparison of the results for soft classification models with conventional pls-da (hard pls-da) showed some benefits including better classification figures of merit.