From 916ac3b69c3ddb55605032b35bcb045894b1ebf9 Mon Sep 17 00:00:00 2001
From: Ross <xkjq@ymail.com>
Date: Mon, 11 May 2026 12:27:28 +0100
Subject: [PATCH] feat: Improve OCR sample selection and detection of clinical
 identifiers

---
 atlas/templates/atlas/new_uploads.html | 69 ++++++++++++++++++++++----
 1 file changed, 60 insertions(+), 9 deletions(-)

diff --git a/atlas/templates/atlas/new_uploads.html b/atlas/templates/atlas/new_uploads.html
index 671dffd5..7f168fc5 100644
--- a/atlas/templates/atlas/new_uploads.html
+++ b/atlas/templates/atlas/new_uploads.html
@@ -60,7 +60,7 @@
                     </div>
                 </div>
                 <div class="form-text mt-2">
-                    OCR now runs sequentially on one sample image per series (capped) to reduce browser memory use.
+                    OCR runs sequentially on multiple sample images per series (capped) to improve burned-in text detection.
                 </div>
             </details>
         </div>
@@ -188,7 +188,8 @@
         };
 
         const SITE_PREFIXES = ["ref", "rk9", "ra9", "rh8", "rbz", "rba"];
-        const OCR_MAX_SERIES_SAMPLES = 40;
+        const OCR_SERIES_SAMPLE_COUNT = 3;
+        const OCR_MAX_SAMPLES = 120;
         const OCR_CANVAS_MAX = 768;
         const DUPLICATE_BATCH_SIZE = 500;
 
@@ -665,7 +666,49 @@
 
         function detectOcrPhi(text) {
             const value = (text || "").toLowerCase();
-            return value.includes("accession") || value.includes("patient") || /(ref|rk9|rh8|ra9|rbz|rba)\d+/i.test(value);
+
+            // Catch common clinical identifier markers and identifier-like tokens.
+            if (value.includes("accession") || value.includes("patient") || value.includes("nhs") || value.includes("hospital")) {
+                return true;
+            }
+
+            return /(ref|rk9|rh8|ra9|rbz|rba)\d+/i.test(value)
+                || /\b\d{3}[\s-]?\d{3}[\s-]?\d{4}\b/.test(value)
+                || /\b\d{10}\b/.test(value)
+                || /\b\d{6,8}\b/.test(value)
+                || /\b(?:dob|d\.o\.b|birth|born)\b/.test(value);
+        }
+
+        function buildOcrSamples(records) {
+            const includeRecords = records.filter((record) => record.include);
+            const seriesMap = {};
+            for (const record of includeRecords) {
+                if (!seriesMap[record.seriesUid]) {
+                    seriesMap[record.seriesUid] = [];
+                }
+                seriesMap[record.seriesUid].push(record);
+            }
+
+            const samples = [];
+            for (const seriesRecords of Object.values(seriesMap)) {
+                if (!seriesRecords.length) continue;
+
+                const candidateIndexes = [
+                    0,
+                    Math.floor((seriesRecords.length - 1) / 2),
+                    seriesRecords.length - 1,
+                ];
+
+                const used = new Set();
+                for (const index of candidateIndexes) {
+                    if (used.has(index)) continue;
+                    used.add(index);
+                    samples.push(seriesRecords[index]);
+                    if (used.size >= OCR_SERIES_SAMPLE_COUNT) break;
+                }
+            }
+
+            return samples.slice(0, OCR_MAX_SAMPLES);
         }
 
         async function getDownscaledOcrBlob(file) {
@@ -715,16 +758,18 @@
                 return;
             }
 
-            const sampleBySeries = {};
-            for (const record of window.uploadPreview.records) {
-                if (!sampleBySeries[record.seriesUid]) sampleBySeries[record.seriesUid] = record;
+            if (!window.Tesseract || typeof window.Tesseract.createWorker !== "function") {
+                window.uploadPreview.records.forEach(r => { r.ocrWarning = false; });
+                toastr.error("OCR engine is unavailable in this browser session.");
+                return;
             }
 
-            const samples = Object.values(sampleBySeries).slice(0, OCR_MAX_SERIES_SAMPLES);
+            const samples = buildOcrSamples(window.uploadPreview.records);
             if (!samples.length) return;
 
             const worker = await ensureOcrWorker();
             let index = 0;
+            const flaggedSeriesUids = new Set();
 
             for (const record of samples) {
                 index += 1;
@@ -733,11 +778,17 @@
                     const blob = await getDownscaledOcrBlob(record.file);
                     const result = await worker.recognize(blob);
                     const text = result?.data?.text || "";
-                    record.ocrWarning = detectOcrPhi(text);
+                    if (detectOcrPhi(text)) {
+                        flaggedSeriesUids.add(record.seriesUid);
+                    }
                 } catch (err) {
-                    record.ocrWarning = false;
+                    // Ignore per-image OCR errors and continue with other samples.
                 }
             }
+
+            window.uploadPreview.records.forEach((record) => {
+                record.ocrWarning = record.include && flaggedSeriesUids.has(record.seriesUid);
+            });
         }
 
         async function calculateBlake3Hash(file) {