WCE-SqueezeNet: Targeting on-capsule CNN inference for energy efficient microcontroller-based endoscopy

Wireless Capsule Endoscopy (WCE) is emerging as an important alternative to traditional diagnostic approaches for gastrointestinal diseases. The advantages in terms of comfort and completeness of the inspection encourage the development of innovative and efficient solutions to address two main issues: the time-consuming analysis of the significant amount of data produced and the tight power budget connected to the battery life. In this work, we adopt the edge-processing paradigm to evaluate a suitable Convolutional Neural Network for the recognition of relevant diseases directly on the capsule, reducing the data stream to frames where signs of relevant diseases are detected, thus improving the efficiency of the capsule while assisting medical professionals in the analysis. The WCE-SqueezeNet model was evaluated on WCE frames from the Kvasir and ETIS-Larib, LD Polyp Video, CVC-ColonDB and CVC-CLinicDB datasets, reaching respectively 98.88%, 90.2%, and 80.4% accuracy. Direct measurements of a parallel implementation on the GAP9 multi-core platform demonstrate classification at a 16 fps rate, with 61 ms inference time and 30.6 mW average core power consumption.