Neural Network Training Data Profoundly Impacts Texture-Based Intravascular Image Segmentation

Segmentation of variably differentiated and low-frequency elements in a complex image is challenging. Improving sensitivity demands often prohibitive decreases in specificity. This is particularly the case in intravascular imaging, where detection of heterogeneously dispersed lesion elements, which are often less evident than normal structures, is essential. Modalities including optical coherence tomography (OCT) provide cross-sectional images of coronary arteries that reveal atherosclerotic plaques. Manual plaque segmentation is time consuming and error prone; automated methods are quicker but dictate accuracy tradeoffs. We developed a neural network-based method for automatic detection of calcified plaques in OCT images using texture-based features and examined how underlying training data distribution impacts sensitivity and predictive value. The method assesses each pixel, rather than a patch, as an independent unit, enabling precise control of training data distribution while simultaneously decreasing reliance on massive imaging datasets for training. Pixels from 30 manually annotated OCT images of calcified plaques were used to train the neural network. Several texture measures were computed for the local neighborhood of each pixel and used as inputs to a multi-layered neural network. The ratio of pixels of each class in the training dataset was then varied and the resulting network performance was compared. Positive predictive value and sensitivity ranged from 0.69 to 0.77 and 0.35 to 0.86, respectively, as the ratio of non-calcified to calcified pixels varied from around 15 to 1, with inverse changes in specificity. The results clearly demonstrate that appropriately balanced data must be carefully curated with thoughtful consideration of the model's application and the clinical imperative being addressed.