The Chinese tongue sole (Cynoglossus semilaevis) is a commercially important marine flatfish in China. Recent advances include whole-genome sequencing and the breeding of a new variety, “Ta You No.1”, traditional manual measurement of growth traits remains labor-intensive, time-consuming, and causes significant stress to the fish, thereby becoming a major bottleneck for further genetic improvement of high-yield strains. To address this limitation, the present study developed an image-based, non-contact phenotyping system for automatic measurement of linear growth traits in Chinese tongue sole using deep learning technology. A total of 1,004 healthy individuals at 14 months of age were examined. Body weight and standardized photographs were collected using an automated device. A two-stage deep learning pipeline was implemented: first, a Mask R-CNN model with a ResNet-50 backbone and feature pyramid network was trained for fish detection and segmentation; second, an RTMPose model with a CSPNeXt backbone and SimCC-based keypoint regression was used for precise localization of ten anatomical key points on each image. From these key points, seven linear traits were calculated: total length (TL), body length (BL), body width (BW), head length (HL), snout length (SL), caudal peduncle length (CPL), and caudal peduncle height (CPH). To validate the accuracy of the automated system, 30 randomly selected fish were measured both manually (using a ruler) and by the machine, and correlation coefficients as well as relative errors were computed. Subsequently, descriptive statistics, Pearson correlation analysis, path analysis, and stepwise multiple regression were performed using SPSS 19.0 to examine the relationships between the linear traits and body weight, and to establish an optimal predictive model for body weight. The validation results demonstrated that the correlation coefficients between machine and manual measurements ranged from 0.75 to 0.96, with TL exhibiting the highest correlation (0.96) and CPH the lowest (0.75). The average relative errors varied from 3.43% (TL) to 11.70% (CPH), indicating that the system provides satisfactory accuracy for most major linear traits, although finer structures such as the caudal peduncle remain more challenging. The coefficients of variation for linear traits ranged from 10.96% to 31.64%, while body weight had the highest coefficient (33.63%), indicating substantial phenotypic diversity within the population, a favorable condition for selective breeding. Pearson correlation analysis showed that all pairwise correlations among the eight traits were highly significant (P<0.01). Body weight was most strongly correlated with BW (r=0.903), followed by HL (0.890), BL (0.835), TL (0.830), SL (0.815), CPL (0.555), and CPH (0.401). The correlations among the five major linear traits (BL, BW, TL, SL, HL) were all above 0.8, revealing strong multicollinearity. Path analysis combined with stepwise regression (entry-and-removal method) retained four linear traits as significant predictors of body weight: BL, BW, HL, and CPL. Their direct path coefficients to body weight were -0.464 (BL), 0.772 (BW), 0.435 (HL), and 0.035 (CPL). Body width exhibited the largest positive direct effect, whereas body length showed a negative direct effect. Substantial indirect effects were also observed; for BW, the total indirect effect via other traits was 1.909, mainly through HL (0.751) and BL (0.731). For all retained traits, the sum of indirect effects exceeded the direct effect, highlighting the complex interplay among linear traits in determining body weight. Determination coefficients further quantified these contributions: the single-trait determination coefficient was highest for BW (0.595), meaning that BW alone explained 59.5% of the variation in body weight. The largest combined determination coefficient was for BW and HL (0.652), confirming these two traits as the most important determinants of body weight. Finally, the optimal multiple regression equation for predicting body weight was established as: TW= 4.076×BW+2.436×HL–0.435×BL+1.352×CPL–185.740. The multiple correlation coefficient R was 0.910, and the coefficient of determination R2 was 0.828, indicating that 82.8% of the variance in body weight could be explained by the four linear traits. The model was highly significant as confirmed by ANOVA. In conclusion, this study successfully established a deep-learning-based, image-driven phenotyping system for non-contact, high-throughput measurement of linear growth traits in Chinese tongue sole. The system demonstrated good accuracy for major traits and meets the practical requirements for breeding applications. Path analysis and determination coefficients revealed that body width and head length are the key direct determinants of body weight, while body length and caudal peduncle length also contribute significantly in combination. The derived regression equation provides a reliable and practical indirect selection criterion for body weight—the most direct but labor-intensive trait—using easily and accurately measurable linear traits. This approach reduces fish handling stress, saves labor, and enables large-scale phenotyping, thereby accelerating the breeding of high-yield varieties. The findings are also valuable for extending similar image-based phenotyping and path analysis methodologies to other flatfish and aquaculture species.