Designing a physically-feasible colour filter to make a camera more colorimetric

Previously, a method has been developed to find the best colour filter for a given camera which results in the new effective camera sensitivities that best meet the Luther condition. That is, the new sensitivities are approximately linearly related to the XYZ colour matching functions. However, with no constraint, the filter derived from this Luther-condition based optimisation can be rather non-smooth and transmit very little light which are impractical for fabrication. In this paper, we extend the Luther-condition filter optimisation method to allow us to incorporate both the smoothness and transmittance bounds of the recovered filter which are key practical concerns. Experiments demonstrate that we can find physically realisable filters which are smooth and reasonably transmissive with which the effective 'camera+filter' becomes significantly more colorimetric.

MON-535 Deep-Machine Learning for Objective Quantification of Nerves in Immunohistochemistry Specimens of Thyroid Cancer

Introduction: Nerves in the cancer microenvironment have prognostic significance, and nerve-cancer crosstalk may contribute to tumour progression, but the role of nerves in thyroid cancer is not known (1). Reproducible techniques to quantify innervation are lacking, with reliance on manual counting or basic single-parameter digital quantification. Aims: To determine if a deep machine learning algorithm could objectively quantify nerves in a digital histological dataset of thyroid cancers immunostained for the specific pan-neuronal marker PGP9.5. Methods: A training dataset of 30 digitised papillary thyroid cancer immunohistochemistry slides were manually screened for PGP9.5 positive nerves, annotated using QuPath (2). 1500 true positive nerves were identified. This dataset was used to train the deep-learning algorithm. First, a colour filter identified pixels positive for PGP9.5 (Model 1). Then, a manually tuned colour filter and clustering method identified Regions of Interest (ROIs): clusters of PGP9.5 positive pixels that may represent nerves (Model 2). These ROIs were classified by the deep learning model (Model 3), based on a Convolutional Neural Network with approximately 2.7 million trainable parameters. The full model was run on a testing dataset of thyroid cancer slides (n=5), containing 7-35 manually identified nerves per slide. Model predictions were validated by human assessment of a random subset of 100 ROIs. The code was written in Python and the model was developed in Keras. Results: Model 2 (colour filter + clustering only) identified median 2247 ROIs per slide (range 349-4748), which included 94% of the manually identified nerves. However, most Model 2 ROIs were false positives (FP) (median 85% FP, range 68-95%), indicating that Model 2 was sensitive but poorly specific for nerve identification. Model 3 (deep learning) identified fewer ROIs per slide (median 1068, range 150-3091), but still correctly identified 94% of manually annotated nerves. Of the additionally detected ROIs in Model 3, median FP rate was 35%. However, in slides where higher non-specific immunostaining was present, then the number of FP ROIs was >90%. Conclusion: Simple image analysis based on colour filtration/cluster analysis does not accurately identify immunohistochemically labelled nerves in thyroid cancers. Addition of deep-learning improves sensitivity with acceptable specificity, and significantly increases the number of true positive nerves detected compared to manual counting. However, the current deep learning model lacks specificity in the setting of non-specific immunostaining, which is a basis for improving further iterations of this model to facilitate study of the significance of innervation of thyroid and other cancers. References: (1) Faulkner et al. Cancer Discovery (2019) doi: 10.1158/2159-8290.CD-18-1398. (2) Bankhead P et al. Sci Rep 2017;7(1):16878.