Transfer Learning Based on DenseNet-۱۲۱ Model: A Deep Learning Approach to Promote Detecting Breast Cancer in Thermogram

In the last decade, the use of thermography has become an attractive research topic in thefield of medical imaging for diagnosing breast cancer, thanks to its harmlessness, cheapness, and earlydetection potential. However, the variety of the thermal patterns related to the cancerous lesion and theirseparation from the normal body temperature profile has made the interpretation of thermograms morecomplicated than the images obtained from other diagnostic modalities. This issue has caused muchattention to be paid to the use of artificial intelligence-based methods, in particular, deep neuralnetworks as a tool to detect the presence of cancerous lesions in breast thermograms. Since deep neuralnetworks are considered as a member of the family of big data methods, their optimal performance inthis field is highly dependent on the availability of a large collection of breast thermograms, which sucha vast database is not yet available due to the young age of breast thermography. In this article, the useof transfer learning concept is proposed and examined as a solution to improve the performance of deepneural networks in breast cancer diagnosis by thermograms.In our approach, the DenseNet-۱۲۱ model which has been pre-trained by ImageNet database, containingthan ۱۴ million images and ۱۰۰۰ classes, is reused as the starting point for a model for distinguishinghealthy and cancerous breast tissues via a fine-tuning scheme. This allows us to address the challenge ofthe large amount of computing and storage resources required to develop an effective breast cancer deeplearning based detector. The testbed for evaluation of the proposed scheme was provided by utilizingTensorflow which is an open-source set of Python machine learning module, with a NVIDIA ۲۰۸۰ TIGPU with ۳۲ GB RAM. Furthermore, Google Colab, a GPU framework made available by Google wereused in order to run the computer code. The tests were performed on DMR-IR dataset which includes۷۶۰ thermal images for the sick and ۷۶۲ thermal images for the healthy class and have been dividedinto three categories of training, validation and test subsets, with ratios of ۶۰%, ۱۵% and ۲۵% amongthe total images respectively. The results of multiple experiments demonstrate the acceptableperformance of the proposed technique in terms of obtaining ۸۴% accuracy, ۷۵% sensitivity, and ۹۳%specificity in distinguishing healthy and cancerous breasts. According to demonstrated results,the use of transfer learning paradigm based on DenseNet-۱۲۱ may be considered as an option withacceptable potential in constructing deep neural models in order to interpret breast cancer thermograms.

کلیدواژه ها:

Breast cancer ، Thermography ، Deep learning ، Transfer learning ، DenseNet-۱۲۱