Abstract
During this collaboration between CILcare and KeenEye Technologies, a full pipeline has been designed to automatically classify and quantify the number of hair cells in 3D cochlea images. This project introduced many challenges with regard to specific pre- and post-data processing and an adaptive model for 3D object detection. The model has been trained using transfer learning with mini batch images keeping the context information around the different types of cells. This new automatic counting method performed 10 times faster than humans, with on average 3.5 min to analyze one fragment image. The algorithm gave performance metrics of 90% for precision and 70% for sensitivity. While the precision value is good, additional work is needed to increase the overall sensitivity and reduce its variance. In addition, an objective quantification method to detect tinnitus on rats was developed in collaboration between CILcare and Charles Coulomb Laboratory (L2C-BioNanoNMRI team). Tinnitus, a phantom auditory sensation, which occurs in the absence of an external sound stimulus, is generated presumably within the auditory brain. Here we focus on the inferior colliculus (IC), a midbrain structure that integrates auditory information from both ears as well as information from other sensory systems. Some studies reveal neural hyperactivity in the IC after salicylate drug administration. In this study, we present an innovative manganese-enhanced magnetic resonance imaging (MEMRI) analysis method called ∆R2/R2. This quantitative method detects 1H NMR relaxation rate changes in the absence or presence of tinnitus. The ∆R2/R2 method generates relevant data comparable to those obtained with the signal-to-noise ratio (SNR) and signal intensity ratio (SIR) methods when manganese is administered by the transtympanic or intraperitoneal route. A major advantage of the ∆R2/R2 method is that it is automatic, robust, and reveals quantitative markers compared to qualitative methods like SNR and SIR.