In this paper, we proposed a reduced-block-Shufflenet (RB-ShuffleNet) for thermal breast cancer detection. RB-ShuffleNet is a modification of Shufflenet obtained by reducing blocks from the original architecture. The images for training and testing were obtained from Database for Mastology Research (DMR). First, we detected and cropped the image based on the region of interest (ROI), in which the ROI is determined by using the red intensity profile. Then, the ROI images were trained using RB-ShuffleNets. In the experiments, we built eight architectures, based on ShuffleNet, each with a different number of reduced blocks. The result showed that RB-Shufflenet with four reduced blocks had fewer than 50% of the learning parameters of the original Shufflenet, without compromising its performance. The RB-ShuffleNet with up to four reduced blocks could achieve 100% testing accuracy. Furthermore, The RB-ShuffleNets performed better than MobileNetV2 and resulted in higher accuracy when fed with ROI images. Due to its light structure and good performance, we recommend RB-ShuffleNet as mobile-based CNN model which is preferable to implement in breast cancer detection.