OCR handles low quality medical scans in two stages. It first cleans the image by straightening the page, removing specks and separating text from background. It then reads what is left. Cleanup fixes many crooked or noisy pages, but it cannot restore detail the scan never captured, so some characters still come out wrong and need checking by a person.
What makes a medical scan low quality for OCR?
A medical scan is low quality for OCR when the image makes characters hard to separate from each other or from the background. The usual causes are skew, noise, low contrast, heavy compression and fax artifacts. Each one damages the shapes OCR relies on in a different way, and many poor scans have more than one.
Skew is a page scanned at an angle. Lines of text run uphill or downhill across the image, so the OCR engine has trouble finding where one line ends and the next begins. A page pulled sideways by a staple or placed crookedly on the glass is the common source.
Noise is stray marks that are not part of the text. The Tesseract OCR documentation on improving output quality defines it as "random variation of brightness or colour in an image, that can make the text of the image more difficult to read." On medical paperwork, noise comes from dust on the scanner glass, speckled recycled paper, and the grain of a photocopy of a photocopy.
Low contrast means the text is not much darker than the paper. Faded thermal printouts, pencil notes, light carbon copies and old dot-matrix reports all fall here. A yellowed page with gray text gives the engine very little difference to work with.
Heavy compression happens after the scan. To save space, some software stores images in formats that throw away fine detail. At strong settings, the edges of letters blur into blocks and halos, and thin strokes such as the crossbar of a 4 or a decimal point can vanish.
Fax artifacts are the marks a fax transmission leaves behind: jagged letter edges, streaks across the page, dropped lines and text that has been forced into pure black and white. A faxed referral that was itself printed from an earlier fax carries the damage of both trips.
Resolution is the other common cause, and it has its own guide on the best scan resolution for medical document OCR. The low-resolution images that arrive by fax, where rescanning is not possible, are covered in the medical faxes and referrals category.
Image cleanup steps OCR applies to poor medical scans
Image cleanup, often called preprocessing, is the work OCR software does on a scan before it tries to read any characters. The three main steps are deskewing, denoising and binarization. Each one prepares the image so the recognition step sees straight lines of dark text on a plain light background.
Deskewing
Deskewing rotates the page image until the lines of text are level. The software measures the angle of the text lines, then turns the image by the opposite amount. The Tesseract OCR guide to improving quality gives the same advice: "rotate the page image so that the text lines are horizontal." Deskewing works well on pages that are tilted by a few degrees. It struggles with pages that are curled or folded, where different parts of the page tilt by different amounts.
Denoising
Denoising removes small specks that are not part of any letter. A common method looks for isolated dots smaller than the thinnest stroke of a character and erases them. The risk on medical paperwork is obvious: a decimal point in a lab value or a dollar amount is also a small isolated dot. Aggressive denoising can remove it along with the dust.
Binarization
Binarization turns a grayscale or color image into pure black and white, so each pixel is either text or background. The software picks a brightness threshold and sorts every pixel against it. On an evenly lit page, this works well. The Tesseract documentation notes that the result "can be suboptimal, particularly if the page background is of uneven darkness," such as a page with a shadow along the spine or a coffee stain across one corner. Methods that set a different threshold for each region of the page handle uneven pages better than a single threshold for the whole image.
Border removal and stroke repair
Two smaller steps often run alongside the main three. Border removal crops away the dark edges a scanner can leave around a page, which the Tesseract documentation warns can be read as extra characters. Stroke repair thins letters that have bled together or thickens letters that have broken apart. The documentation mentions an erosion technique for "heavy ink bleeding", which shrinks each dark stroke slightly so neighboring letters separate.
How do low quality medical scans show up in OCR output?
Low quality medical scans show up in OCR output as wrong characters, missing characters and text in the wrong place. The output still looks like normal text, which is what makes these errors dangerous. A misread member ID looks just as confident on screen as a correct one unless the software reports its doubt.
The character confusions that matter most on medical paperwork are the ones that change a value without making it look broken:
- Lost or added decimal points. A lab result of 1.5 read as 15, or a charge of $45.00 read as $4500. Noise removal and fax streaks both cause this.
- Look-alike characters. 0 and O, 1 and l and I, 5 and S, 8 and B, 6 and G. A member ID such as the fictional "MB0018-5Z (example)" can come out with one letter swapped for a digit and still look plausible.
- Merged and split characters. "rn" read as "m", or a bled-together "11" read as "U". Low contrast and ink bleed are the usual causes.
- Dropped minus signs and slashes. A date of 03/04/2026 read as 0304/2026, or a negative adjustment that loses its sign.
- Merged lines and columns. On a skewed table, a value from one row lands in the next row's line, so an amount ends up attached to the wrong service.
Scan defects, the errors they cause and the fix
The table below links each defect to its typical OCR error and the fix that usually works. All values in the examples are fictional.
| Scan defect | Typical OCR error | Example (fictional) | Fix |
|---|---|---|---|
| Skew of a few degrees | Text lines merge, table rows shift | Allowed amount attached to the wrong service line | Cleanup (deskew) |
| Curled or folded page | Uneven lines, words dropped at the fold | Half of a diagnosis description missing | Rescan flat |
| Dust or specks | Extra punctuation, stray characters | Charge of $120.00 read as $1,20.00 | Cleanup (denoise), then clean the glass |
| Faint pencil or thermal print | Characters missing entirely | A handwritten refill count read as blank | Rescan in grayscale |
| Shadow or stain on part of the page | Region lost after binarization | Lower third of a lab table empty | Cleanup (regional threshold) or rescan |
| Heavy compression | Thin strokes vanish, letters blur | Decimal point lost: 2.5 read as 25 | Rescan with lighter compression |
| Fax streaks and jagged edges | Look-alike swaps, merged characters | Member ID 0 read as O | Human review |
| Handwriting on a poor copy | Words misread or unreadable | Prescriber name garbled | Human review |
Handwriting deserves its own note. A poor scan makes handwriting harder still, but even a clean scan of handwriting is a different problem from printed text. The post on whether OCR can read handwritten medical forms explains why.
When to rescan a medical document instead of fixing it in OCR
Rescan a medical document instead of fixing it in OCR when the defect removed information rather than hid it. Cleanup can straighten, sharpen and separate what the scan captured. It cannot invent strokes that were never recorded. If you cannot read a value yourself on the scanned image, software is unlikely to read it correctly.
A simple test separates the two cases. Zoom in on the problem area of the scan. If the characters are there but crooked, speckled or gray, cleanup has something to work with. If the characters are broken, missing, blurred into blocks or cut off at the edge of the image, only a new scan will help.
Rescan in these cases:
- A page is cut off, folded over or partly covered by a sticky note.
- Text is faint because the scan used black and white mode. A grayscale rescan often brings it back.
- A shadow or glare covers part of the page.
- Compression has left blocky halos around letters.
- The page is missing, doubled or upside down in a way rotation did not fix.
Rescanning is only possible while you still have the paper. That is the strongest reason to check every batch before the originals are boxed or destroyed. For pages you cannot rescan, such as an incoming fax or a copy received from another provider, ask the sender for a cleaner copy when a field matters, and send the rest through review.
Teams deciding whether to build their own cleanup and review pipeline or buy one face this trade-off at scale. The post on build vs buy for medical document extraction covers what each path asks of a team.
Routing unreadable medical scan fields to a person
Routing unreadable medical scan fields to a person means the software sends any field it cannot read with enough certainty to a reviewer, instead of passing it straight into a billing or records system. The trigger is a confidence score: a number the software assigns to each field to show how sure it is of the reading.
The process works in three parts. The software reads every field and gives each one a confidence score. The team sets a threshold. Any field under that threshold goes into a review queue, where a person compares it with the original page and approves or corrects it. Fields above the threshold go through without a stop.
Field-level routing is what makes poor scans manageable. A smudged fax may have twenty fields, of which eighteen read cleanly and two do not. Sending the whole page back to a person wastes their time on the eighteen. Sending only the two, with the spot on the page highlighted, keeps the review short.
Schema Care scores every field it reads, holds fields below your threshold for a reviewer who sees each one highlighted on the source page, and logs every approval and correction with who made it and when. You can see the threshold at work on a fictional EOB in the homepage demo.
Choosing the threshold, calibrating scores and staffing the queue are topics of their own. The human review category covers confidence scores, thresholds and review workflows in depth.