CNN Architecture for Surgical Image Segmentation Systems with Recursive Network Structure to Mitigate Overfitting
Taito Manabe, Koki Tomonaga, Yuichiro Shibata · 2019
Laparoscopic surgery, a minimally invasive camera-aided surgery, is performed commonly. However, it requires a camera assistant who holds and maneuvers a laparoscope. If the laparoscope can be controlled automatically using a robot, a surgeon can perform the operation without a camera assistant, which would be beneficial in the areas suffering from lack of surgeons. In this paper, a prototype image segmentation architecture, based on a convolutional neural network, is proposed to realize an automatic laparoscope control for cholecystectomy. Since the learning dataset is annotated manually by a few surgeons, its scale is currently quite limited. Therefore, we devised a recursive network structure, with some sub-networks which are used multiple times, to mitigate overfitting. Furthermore, instead of the common transposed convolution, the flip-based subpixel reconstruction is introduced into upsampling layers. Evaluation results reveal that these improvements bring better classification accuracy without increasing the number of parameters. The system shows a throughput sufficient for real-time laparoscope robot control with a single NVIDIA GTX 1080 GPU.