DAFA: A Dual-Awareness Feature Aggregator for Table Structure Recognition on Medical Examination Reports

Table structure recognition (TSR) is crucial for document analysis, particularly for medical examination report tables (MERTs), impacting efficiency and decision-making in healthcare. Most models for TSR utilize either Graph Convolution Neural Networks (GCNs) or Transformers with html sequences for structure recognition. These methods, however, face challenges with graph inductive bias and instability in training, respectively. We observe that cells within the same row or column of a table not only are closely aligned in their vertical and horizontal coordinates, respectively, but also exhibit highly similar features. In previous work, the spatial feature of coordinates was often only used for concatenation with image features, text features, etc. We believe that explicitly utilizing the unique spatial properties of tables can better encode table features. In this paper, we introduce a novel structure named Dual-Awareness Feature Aggregator (DAFA) for table, which leverages attention mechanisms to effectively extract table features. Based on it, we design an end-to-end model called DAFA-Net requiring only images as input, without the need for additional information such as texts. In addition, we try to address the prevalent challenge of recognizing cross-row and cross-column cells in TSR — a scenario frequently encountered in medical examination reports — by introducing a modified focal loss known as CRCC loss. We conduct extensive experiments on four popular datasets. This includes a dataset specifically dedicated to medical data and others that mirror the complexity typically encountered in medical tables. Experimental results show the effectiveness and potential of our DAFA-Net for TSR within the healthcare sector.