Traditional image-based detection methods for liver cancer have problems of large overlap error and low accuracy; thus, a paradigm based on the overlapping error for image detection has been proposed in previous works. In addition, biochemical sensing systems, such as lab-on-a-chip, BioMEMS/NEMS, and biomimetic systems, have stimulated much interest in the research community. We propose an adaptive region of interest detection method based on a convolutional neural network. Deep learning is carried out for some layers of the convolutional neural network, and parameters are optimized by using the improved loss function. Image features are enhanced and extracted in combination with histogram equalization, liver cancer regions of interest are extracted on the basis of an extensible markup language file, and the adaptive detection of liver cancer shadow-related areas is completed via computing through online detection and annotating a sequence of computed tomography images. Experimental results show that the proposed algorithm can effectively reduce the overlap error and improve the detection accuracy. When the number of image sets was 300, the detection accuracy of this method was 95.5%.