Mucositis, an inflammatory condition affecting the mucosa, is a prevalent and debilitating complication of cancer treatments, commonly encountered in head and neck cancer (HNC) patients undergoing radiotherapy or chemoradiotherapy. Despite being clinically visible and representing an intense inflammatory response, accurate prediction of mucositis at early stage is crucial for effective treatment and improved quality of life. We proposed the light-weight convolutional neural network (Lw-CNN) framework for automated binary-class classification of thermograms into mucositis grades, including GRADE 0 corresponds to the absence of mucositis, GRADE I indicates either asymptomatic or mild symptoms enabling early identification of patients at risk. Our study utilized thermogram data from 50 HNC patients over a 4-week period at Homi Bhabha Cancer Hospital in India, captured using a FLIR E-60 device. Using the ADAM optimizer resulted in 79% accuracy and the RMS Prop optimizer achieved 74% accuracy, which was also validated by the senior oncologists. In future, this low-cost thermography approach holds promise for widespread implementation in large-scale cancer patient populations, enabling early prediction of mucositis of all grades during the course of treatment.