Post-weld evaluation does not readily capture the dynamic evolution of nugget formation in resistance spot welding. In this study, an embedded ultrasonic probe was used to acquire A-scan signals during resistance spot welding of stainless steel. Consecutive signals were assembled into M-type ultrasonic images to analyze echo changes across the thermal cycle and under different heat inputs. Using 12 specimens with metallographic measurements of both nugget diameter D and thickness H, five- and three-dimensional acoustic features were extracted for D and H prediction, respectively. Random forest (RF) regression models were established and evaluated using leave-one-out cross-validation (LOOCV). The M-type images showed the streak evolution of interface echoes, back-wall echoes, and nugget-related echoes, reflecting differences in acoustic responses across welding stages and nugget states. The coefficient of determination (R²), mean absolute error (MAE), and root mean square error (RMSE) were 0.588, 0.674 mm, and 0.798 mm for D, respectively, with corresponding values of 0.570, 0.150 mm, and 0.198 mm for H. These preliminary results indicate quantitative correlations between the extracted low-dimensional acoustic features and nugget geometry, providing an experimental basis for nugget geometry evaluation using in-process ultrasound.