Enabling Bias-Dependent Electronic Morphology Analysis of Single Molecules in STM via Deep Segmentation with Noise-Aware Calibration

Scanning tunneling microscopy (STM) images are frequently affected by low-frequency vibrations, substrate-induced background variations, and bias-dependent contrast changes, which degrade molecular feature responses and hinder reliable segmentation. To address this challenge, we develop an STM-oriented Feature Pyramid Network (FPN) + Dual-Path Intensity Calibration (DPIC) framework by adapting a DPIC module, originally derived from a cloud-noise calibration mechanism, to the specific characteristics of STM molecular images. In this framework, DPIC is reformulated as a noise-aware feature calibration module that suppresses low-response background interference while preserving foreground molecular contours. We integrated DPIC into the FPN architecture and systematically evaluated the model on an in-house STM molecular image dataset. Expanded baseline comparisons and image-level five-fold cross-validation were further conducted to assess segmentation performance and stability. The proposed FPN + DPIC model achieved an Intersection over Union (IoU) of 0.8347 ± 0.0073 and a Dice score of 0.9077 ± 0.0050, outperforming representative deep learning, traditional segmentation, and scanning tunneling microscopy (STM)-related comparison methods under the same evaluation protocol. Furthermore, we applied the model to bias-dependent STM images of single 4,4^′-Methylenebis(N,N-diphenylaniline) (MTDATA) molecules. By introducing Radial Skewness as a contour-derived geometric descriptor, we quantified the bias-dependent evolution of molecular electronic-state morphology. Statistical correlation analysis further showed that Radial Skewness is negatively associated with the Highest Occupied Molecular Orbital (HOMO)-related spectral response within the measured STM/scanning tunneling spectroscopy (STS) series. This work provides a noise-aware segmentation strategy for the tested MTDATA STM molecular images and demonstrates the potential of segmentation-assisted morphology analysis for quantitative interpretation of bias-dependent STM/STS data under the investigated molecular system and imaging conditions.