Can the Azure AI Document Intelligence Prebuilt Layout Model Be Custom Trained for Better Semantic Table Understanding?

Nazmaheptulla Thogaragunta 0 Reputation points
2026-07-03T08:02:06.7733333+00:00

Hi ,

We are currently using the Azure AI Document Intelligence Prebuilt Layout model to extract tables from healthcare Itemized Billing (IB) documents.

The model performs well for detecting tables, but we are facing challenges with column boundary identification in complex billing tables. Since the Layout model primarily relies on the visual structure of the document, it sometimes assigns values to incorrect columns when there are inconsistent layouts, merged cells, missing borders, or varying table formats.

Our documents typically contain columns such as:

  • Date of Service

HCPCS/CPT Code

Revenue Code

Description

Units

Charge Amount

Total Amount

Although these fields have recognizable semantic meanings, the model does not appear to use this information when determining column boundaries.

We would like to understand the following:

Is it possible to custom train or fine-tune the Prebuilt Layout model so that it better understands the semantics of healthcare itemized billing documents?

If the Prebuilt Layout model cannot be customized, is there another Azure AI Document Intelligence model that supports this type of semantic understanding for table extraction?

Are there any recommended approaches or best practices for improving column boundary detection in complex tables without relying heavily on post-processing?

Is semantic-aware table extraction or domain-specific training for the Layout model on Microsoft's roadmap?

Our goal is to improve the accuracy of extracting structured data from healthcare itemized billing documents while minimizing custom post-processing logic.

Any guidance or recommendations would be greatly appreciated.

Thank you.

Azure Document Intelligence in Foundry Tools
0 comments

1 answer

Sort by: Most helpful
Most helpful Newest Oldest
  1. Thanmayi Godithi 10,990 Reputation points Microsoft External Staff Moderator
    2026-07-03T10:09:44.43+00:00

    Hi Nazmaheptulla Thogaragunta ,

    Great questions — here's the direct rundown:

    1. Can the Prebuilt Layout model be fine-tuned/custom-trained? No. Document Intelligence doesn't allow training or fine-tuning on top of any prebuilt model (Layout included). Prebuilt models are fixed; customization always means building a standalone custom model from your own labeled documents. Layout is also purely visual-structure-based, so it doesn't use the semantic meaning of your columns when deciding boundaries — which is exactly what you're hitting on merged cells / missing borders / variable formats.
    2. Is there a model that adds semantic understanding for tables? Yes — a Custom Neural (custom extraction) model. Unlike Layout, it combines layout and language features, so it learns the meaning of each field and handles structured/semi-structured documents with variable layouts. You label your billing table once (with the 7 columns — Date of Service, HCPCS/CPT, Revenue Code, Description, Units, Charge Amount, Total Amount) on as few as ~5 examples and train. This is the recommended path for your itemized-billing scenario.
    3. Best practices to improve column-boundary accuracy (minimal post-processing):
    • Switch the table extraction from Layout → Custom Neural, labeling the line items as a structured table field with your named columns so the model learns column semantics rather than relying on borders.
    • Train on a representative variety of IB formats (merged cells, missing borders, different providers) so it generalizes.
    • If you have several distinct IB templates, add a Custom Classifier to route each format to the right model (or use a composed model).
    • Use the v4.0 (2024-11-30 GA) neural model, which adds table/row/cell-level confidence — gate only the low-confidence cells for human review instead of broad post-processing.
    • Short-term, you can also try Layout with features=keyValuePairs or query fields for specific named values, but for tabular billing lines the custom neural table is the durable fix.
    1. Semantic-aware Layout / domain-specific Layout training on the roadmap? I don't have a public roadmap commitment to share on that; the supported approach today is the custom neural model above.

    Docs:

    Kindly let us know if the above helps or you need further assistance on this issue.

    If the answer is helpful, please click "Accept Answer" and kindly upvote it. If you have extra questions about this answer, please click "Comment".

    Was this answer helpful?

    0 comments

Your answer

Answers can be marked as 'Accepted' by the question author and 'Recommended' by moderators, which helps users know the answer solved the author's problem.