Assessing the expression of the Ki-67 protein and tumor-infiltrating lymphocytes (TILs) in histopathological images plays a crucial role in cancer prognosis. However, accurately identifying these features—particularly TILs—remains a significant challenge for current deep learning methods due to morphological variability and signal inconsistency across samples. A common limitation is the dependence on manually defined or sub-optimally tuned classification thresholds and class weights, which often hinder the model’s ability to focus effectively on difficult channels. In this study, we propose an enhanced approach comprising two key components: (1) an automated threshold optimization strategy based on a multi-initial heuristic search to eliminate dependence on fixed thresholds, and (2) a dynamic weight adjustment mechanism that enables the model to adaptively prioritize challenging signal channels such as TILs during training. Experimental results on histopathological datasets demonstrate that our method significantly improves recognition performance, with average F1-scores for TILs, Ki-67-positive, and Ki-67-negative cells increasing by 0.15% compared to baseline methods using static thresholds and weights.