<p>This study provides the spectrometric features of mud crabs (Scylla paramamosain) extracted under both in-vitro and semi-in-vivo conditions aiming at automatic internal quality grading based on machine-learning technique. These internal qualities, such as meat yield and ovarian fullness, are currently assessed through subjective manual methods, and its automation remains challenging because of the lack of sufficient datasets and promising features to train machine-learning models. Our previous study demonstrated the potential of spectrometric features for nondestructive discrimination of crab components (meat, ovary, liver, shell) using transmittance spectra of the component suspensions under in-vitro conditions. However, these did not fully represent intact crab tissues. In this study, we developed a novel optical system to acquire the interactance spectra directly from an intact crab portion under semi-in-vivo condition. The interactance spectra were collected from the 22 half-portions of 11 crabs under semi-in-vivo conditions, together with the corresponding transmission spectra under in-vitro conditions. The results of the Principal Component Analysis revealed substantially improved tissue separation of the interactance-spectral dataset compared with the transmittance-spectral dataset, with the increase of Silhouette Score from 0.578 to 0.762, despite the complexity of the interactance-mode optics. Their Linear Discriminant Analysis achieved excellent Silhouette Scores of 0.985 and 0.970 for interactance and transmittance datasets, respectively, validating the discrimination capability of the spectrometric features. These findings demonstrate the clear utility of the spectrometric features extracted with the proposed interactance-mode optical system, and promote the development of machine-learning based quality grading models.</p>