An automated image analysis pipeline for wide-field optical redox imaging of patient-derived cancer organoids

16 Wide-field optical redox imaging provides a fast and accessible method to monitor metabolic 17 changes in cells and has recently been developed for drug screening in patient-derived cancer 18 organoids (PDCOs). However, manual analysis of wide-field optical redox images is inefficient 19 and laborious for large-scale drug screens. Here, we developed an automated pipeline for PDCO 20 segmentation, single-PDCO tracking, and background correction in autofluorescence images. 21 This pipeline was tested on two imaging systems over a 3-day time-course with two drug doses 22 to demonstrate generalizability across imaging systems. Segmentation was performed using a 23 fine-tuned Cellpose model, which when compared to manual masks, achieved mean Dice scores 24 >0.8 across systems, indicating high reproducibility. Automated single-PDCO tracking was 25 compared to manual tracking and the accuracy of the tracking algorithm exceeded 94% by two 26 metrics, recall and Jaccard index. For background correction, the automated pipeline uses the full 27 field-of-view to reduce sampling bias. Compared to the manual analysis pipeline, the automated 28 pipeline resolves single-PDCO responses with comparable sensitivity to drug treatment but with 29 over 127× faster processing time. This novel automated image analysis pipeline improves 30 throughput and robustness in PDCO image analysis, which increases the accessibility and 31 scalability of wide-field optical redox imaging for PDCO drug screening. 32 33 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint

34 35 Colorectal cancer (CRC) is responsible for more than 50,000 deaths in the US annually1. CRC is 36 a genetically and metabolic heterogeneous disease, but current treatments cannot effectively 37 address patient and tumor heterogeneity2. There is differential sensitivity between patients. Within 38 the same patient, tumors also respond to drugs differently. In addition to genetic heterogeneity, 39 metabolic heterogeneity is apparent in CRC and contributes to drug resistance3. This emphasizes 40 the need to develop a widely accessible and high-throughput drug screening method to 41 understand the relationship between CRC heterogeneity and treatment response. 42 43 PDCOs are three-dimensional in vitro models derived from patient tumors that provide an 44 excellent method to study subclonal resistance and metabolic heterogeneity4–6. PDCOs capture 45 not only driver alterations but also subclones from the original tumor that drive secondary or 46 acquired resistance6. PDCO metabolism provides an early measurement of treatment response 47 compared to PDCO size, cell proliferation, and cell death7–10. However, existing PDCO metabolic 48 assays have limitations in evaluating single PDCOs, longitudinal tracking in time course studies, 49 and maintaining organoid integrity11–20. Therefore, a method that can provide repeated 50 measurements of single PDCO metabolism without altering the PDCO culture could provide 51 valuable insight into CRC treatments that address tumor heterogeneity. 52 53 Optical redox imaging addresses these limitations by monitoring single-PDCO metabolic 54 heterogeneity over time without compromising the organoids with labeling or destructive 55 techniques. Optical redox imaging leverages the autofluorescent properties of metabolic 56 coenzymes nicotinamide adenine dinucleotide (phosphate) and flavin adenine dinucleotide 57 (NAD(P)H and FAD) to measure the optical redox ratio (ORR), or the oxidation-reduction state of 58 a cell. Here, NADH and NADPH are collectively referred to as NAD(P)H due to their overlapping 59 spectral properties21. The ORR of PDCOs is often measured with multi-photon microscopy to 60 resolve single cells, but multiphoton microscopy is costly, requires lengthy imaging times, and is 61 therefore impractical for high-throughput drug screening22. Alternatively, wide-field microscopes 62 are more affordable and accessible than multiphoton microscopes, and optical redox imaging with 63 widefield microscopy only requires standard components such as a broadband excitation source, 64 scientific monochrome camera, and standard 4′,6-diamidino-2-phenylindole (DAPI) and 65 fluorescein isothiocyanate (FITC) filter cubes to collect NAD(P)H and FAD autofluorescence, 66 respectively. Our group and others have used wide-field optical redox imaging at both cellular and 67 tissue levels to evaluate oxidative stress, mitochondrial function, and drug responses23–25. Our 68 prior work has demonstrated that wide-field optical redox imaging is a useful and nondestructive 69 tool to assess patient treatment response by monitoring longitudinal metabolic changes in PDCOs 70 at the single-PDCO level26,27. However, analysis of widefield optical redox imaging currently relies 71 on a manual image analysis pipeline that is labor intensive, time consuming, and can be affected 72 by human fatigue. 73 74 Here, we develop and optimize an automated image analysis pipeline for widefield optical redox 75 imaging, which greatly accelerates drug screens of PDCOs while further improving the 76 accessibility of this technique. To develop this automated pipeline, we first addressed challenges 77 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint with automated PDCO segmentation including variability in PDCO shape, size, and focal plane 78 positioning. Although prior studies have developed open source software to segment PDCOs, 79 mostly based on bright field images, these solutions are not transferrable to autofluorescence 80 images due to differences in contrast and signal to noise ratio (SNR)28–31. Second, PDCO 81 variability requires tracking treatment response over time within the same PDCO to increase 82 sensitivity to drug treatment compared to pooled PDCO analysis26. The automated PDCO tracking 83 technique greatly improves the analysis time and reliability over manual PDCO tracking. Third, 84 the current approach to calibrate each optical redox image requires a manual background region-85 of-interest (ROI) selection; whereas the automated background selection removes this additional 86 manual step while improving the repeatability of the background measurement. Overall, we have 87 developed automated modules for PDCO segmentation, tracking, and optical redox imaging 88

normalization to complete a fully automated image analysis pipeline for wide-field 89 optical redox imaging that improves the throughput, accessibility, and reliability of PDCO drug 90 response measurements. We demonstrate this pipeline across two separate imaging systems to 91 highlight the generalizability of this approach, using a 3-day imaging time-course of PDCOs in 92 response to two doses of drug treatment. 93 94 95 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint

96 97 CRC PDCO isolation and culture 98 All work was conducted with Institutional Review Board (IRB) approval, and informed written 99 consent was obtained from the patient through the University of Wisconsin (UW) Molecular Tumor 100 Board Registry (UW IRB#2015-1370) or the UW Translational Science BioCore (UW IRB#2016-101 0934). All methods were performed in accordance with the relevant guidelines and regulations. 102 CRC tissue was obtained via primary resection and processed as previously described32,33. 103 Processed PDCO suspension was immediately mixed in a 1:1 ratio with Matrigel (Corning) and 104 conditioned media. Droplet suspensions were plated, set for 3-5 minutes at 37°C, and then 105 inverted for at least 30 minutes to solidify the matrix and avoid direct contact of PDCOs with the 106 interface. Plated cultures were overlaid with 450 µL of media, consisting of DMEM/F12 107 (Invitrogen) supplemented with 1x Glutamax (Invitrogen), 10 mM HEPES (Fisher), 50 IU/mL 108 penicillin-streptomycin (Invitrogen), 50 ng/mL EGF (Invitrogen), and mixed 1:1 with WNT3a-109 conditioned media33. PDCOs were incubated at 37°C in 5% CO₂ and media was replaced every 110 48–72 hours. A single CRC PDCO line was used for this study. 111 Romidepsin treatment experiment 112 PDCOs were collected from 24-well culture plates and transferred to 24-well glass bottom plates 113 (Corning), then rested for 24 hours. Pre-treatment (day 0) images of the PDCOs were acquired 114 using a Nikon Ti-S inverted microscope or Keyence BZ-X810 microscope as described below. 115 Romidepsin (MedChem Express; HY-15149) was prepared at 10 mM in DMSO and diluted to 30 116 or 100 nM in culture media, while no DMSO was added to the control media. PDCOs were treated 117 with control media, 30 nM romidepsin, or 100 nM romidepsin for 48 hours, and imaged again at 118 24 hours (day 1) and 48 hours (day 2). 119 Wide-field optical redox imaging 120 This study performs wide-field optical redox imaging with two imaging systems (Nikon and 121 Keyence) to demonstrate the generalizability of our automated image analysis pipeline. 122 Nikon 123 A Nikon Ti-S microscope coupled with a SOLA light engine (300 to 650 nm, Lumencor, Beaverton, 124 Oregon, United States) was used. Images were acquired using NIS Elements software (Nikon), 125 a 4x air, 0.13 NA objective (Plan Apo, Nikon), and a Hamamatsu Flash4 digital CMOS camera 126 (Hamamatsu City, Japan). NAD(P)H was excited for 3 seconds through a 360/40 nm filter at 40% 127 power (0.25 mJ/cm²), and emission was collected using a 400 nm dichroic mirror and a standard 128 4′,6-diamidino-2-phenylindole (DAPI, 460/50 nm) filter. FAD was sequentially excited for 3.5 129 seconds through a 480/30 nm filter at 40% power (0.531 mJ/cm²), with emission collected using 130 a 505 nm dichroic mirror and a standard fluorescein isothiocyanate (FITC, 535/20 nm) filter. Four 131 fields of view (FOV) (2048 x 2048 pixels; 1.61 μm/pixel) of each PDCO preparation treated with 132 control media, 30 nM romidepsin, or 100 nM romidepsin were imaged pre-treatment (day 0) and 133 post-treatment at 24 hours (day 1) and 48 hours (day 2). 134 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint Keyence 135 A Keyence BZ-X810 microscope (Osaka, Japan) coupled with a 40 W LED (Osaka, Japan) using 136 a 4x air, 0.20 NA objective was used. Images were acquired with a 2/3 inch, 2.83 million pixel 137 monochrome CCD (Osaka, Japan). NAD(P)H was excited for 3 seconds at 20% light power 138 through a DAPI filter cube (360/40ex, 400nm dichroic, 460/50em; Keyence BZ-X DAPI filter). FAD 139 was sequentially excited for 3.5 seconds at 40% light power through a custom filter (480/40ex, 140 510 dichroic, 535/50em; Keyence BZ-X FITC filter). Each PDCO population treated with control 141 media, 30 nM romidepsin, or 100 nM romidepsin was imaged pre-treatment (day 0) and post-142 treatment at 24 hours (day 1) and 48 hours (day 2). The same wells of PDCOs were imaged with 143 both the Keyence and Nikon systems, however, on the Keyence system automatic stitching was 144 performed with a 30% overlap between adjacent FOV, yielding one large FOV (3594 x 3486 145 pixels; 2.4 μm/pixel) per well. 146 147 Manual Image Analysis Pipeline 148 Manual segmentation and leading-edge analysis 149 PDCO masks were manually segmented from NAD(P)H intensity images using a CellProfiler 150 pipeline27. PDCOs consist of (from innermost to outermost) necrotic cores, quiescent zones, and 151 proliferating zones. Proliferating zones at the outer edge of PDCOs are the most metabolically 152 active and most sensitive to treatment34,35. To isolate redox changes in the proliferating zone, our 153 group developed “leading-edge analysis” in a recent study36. Briefly, we calculate the radius of 154 each single-PDCO mask, and subtract 20 pixels in Nikon or 13 pixels in Keyence (due to 155 differences in resolution), roughly 32 µm, from the radius to generate a core mask, which is then 156 subtracted from the single-PDCO mask to create a leading-edge mask. The normalized ORR and 157 redox ratio changes were then calculated from leading-edge masks. 158 Sampled Background Normalization of ORR 159 160 To address issues such as variations in different imaging systems, instrumental drift within an 161 imaging system, uneven illumination, and autofluorescence from Matrigel, ORR values were 162 normalized to a background value within each image. In the manual pipeline, background value 163 refers to the average fluorescence intensity of five PDCO-free regions selected by an individual 164 (Fig. 1). 165 The background value is then used to calculate the normalized ORR for each PDCO, defined as 166 in Eq. 1-3: 167 𝐼𝑁𝐴𝐷ሺ𝑃ሻ𝐻 = 𝐼𝑃𝐷𝐶𝑂_𝑁𝐴𝐷ሺ𝑃ሻ𝐻 𝐼𝑏𝑎𝑐𝑘𝑔𝑟𝑜𝑢𝑛𝑑_𝑁𝐴𝐷ሺ𝑃ሻ𝐻 (Eq. 1) 168 𝐼𝐹𝐴𝐷 = 𝐼𝑃𝐷𝐶𝑂_𝐹𝐴𝐷 𝐼𝑏𝑎𝑐𝑘𝑔𝑟𝑜𝑢𝑛𝑑_𝐹𝐴𝐷 (Eq. 2) 169 𝑁𝑜𝑟𝑚𝑎𝑙𝑖𝑧𝑒𝑑 𝑂𝑅𝑅 = 𝐼𝑁𝐴𝐷ሺ𝑃ሻ𝐻 𝐼𝑁𝐴𝐷ሺ𝑃ሻ𝐻+𝐼𝐹𝐴𝐷 (Eq. 3) 170 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint Where IPDCO_NAD(P)H is the mean intensity of NAD(P)H across the leading-edge mask of the PDCO, 171 and Ibackground_NAD(P)H is the mean intensity of the five PDCO-free regions across the entire NAD(P)H 172 image. IPDCO_FAD is the mean intensity of FAD across the leading-edge mask of the PDCO, and 173 Ibackground_FAD is the mean intensity of the five PDCO-free regions across the entire FAD image. 174 For the manual pipeline, ∆ORR is defined as the normalized ORR of an individual PDCO (Eq. 3) 175 minus the average of all PDCOs in that condition measured on day 0 as in Eq 4: 176 ∆𝑂𝑅𝑅𝑝𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛 = 𝑁𝑜𝑟𝑚𝑎𝑙𝑖𝑧𝑒𝑑 𝑂𝑅𝑅𝑜𝑟𝑔𝑎𝑛𝑜𝑖𝑑 𝑑𝑎𝑦 𝑁 𝑜𝑟 ሺ𝑁+1ሻ − 𝑁𝑜𝑟𝑚𝑎𝑙𝑖𝑧𝑒𝑑 𝑂𝑅𝑅𝑎𝑣𝑒𝑟𝑎𝑔𝑒 𝑑𝑎𝑦 ሺ𝑁−1ሻ (Eq. 4) 177 For generalizability, we denote consecutive days as day (N-1), day N, and day (N+1). In this study, 178 they correspond to day 0, day 1, and day 2, respectively. Where “Normalized ORRorganoid day N” is 179 the background normalized ORR of individual PDCO on day 1, and “Normalized ORRorganoid day 180 (N+1)” is the background normalized ORR of individual PDCO on day 2. “Normalized ORRaverage day 181 (N-1)” is the background normalized ORR of the same PDCO population measured on day 0 prior 182 to treatment. 183 184 Automated Image Analysis Pipeline 185 All scripts in the automated image analysis pipeline were written in Python 3.10.12. 186 187 Algorithm overview: PDCO segmentation, longitudinal single PDCO tracking, and 188

correction 189 190 Prior work has demonstrated that wide-field optical redox imaging is a rapid, sensitive, and non-191 invasive method to measure treatment response and heterogeneity in PDCOs26,36,37. To meet the 192 demands of analyzing large-scale drug screens, a more accessible and high-throughput image 193 analysis method is needed. This study aims to address this need by automating PDCO 194 segmentation, longitudinal single PDCO tracking, and background correction. A complete 195 workflow of the improved analysis pipeline is demonstrated here (Fig. 2A-F). 196 197 Automated Segmentation 198 Following the instructions described in "Cellpose 2.0: How to Train Your Own Model," the cyto3 199 model was fine-tuned using the following training parameters: pretrained_model cyto3, n_epochs 200 100, learning_rate 0.1, and weight_decay 0.000138,39. 55 PDCO NAD(P)H intensity images and 201 their manual masks were used to train the model. 15 NAD(P)H intensity images and their manual 202 masks were tested, including 3 images of 30nM romidepsin treatment taken with the Nikon 203 microscope. All images in the training and testing sets were taken with the Nikon microscope. 204 Only whole PDCOs were masked in all of the images used for training and testing. 205 The fine-tuned Cellpose model was used to generate PDCO masks with a batch processing script. 206 PDCOs with pixels touching the borders of the masks were removed by a post-processing script. 207 Only whole PDCO masks were used for the tracking algorithm (Fig. 3) and background mask 208 identification (Fig. 4) described below. 209 210 Tracking Algorithm 211 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint The tracking algorithm (Fig. 3) used the following modules and libraries: glob, matplotlib, tifffile, 212 opencv-python, SciPy, and scikit-image40–42. Here, a mask refers to a binary image of segmented 213 PDCOs, and an object refers to an individual PDCO in the mask. Up to 3 masks (i.e., time points) 214 can be input into the tracking algorithm but the program may be modified to accommodate more 215 masks. Here, the workflow up until the ID consolidation step is illustrated with only the day (N-1) 216 and day N image pair. The same workflow is repeated for the day N and day (N+1) image pair 217 (Fig. 3A-C). After ID consolidation, IDs obtained from the day N and day (N+1) pair are 218 propagated to day (N-1) (Fig. 3D-E). 219 First, the tracking algorithm preprocesses the masks by removing small objects below 1000 pixels 220 (Fig. 3A-B). Next, within an image time-series, day (N-1) mask is coarsely registered to day N 221 mask. 2D cross-correlation scores are computed within an offset search area of ±25% of the mask 222 size in X and Y directions. The translation vector resulting in the highest cross-correlation score 223 between the two masks is identified. This vector is applied to coarsely register day (N-1) mask to 224 day N mask within an image time-series. 225 To match each PDCO over time, objects in the day (N-1) mask are sorted by size to prioritize 226 larger objects as they tend to have more prominent features. Each object in the day (N-1) mask 227 is extracted as a template, and a search region (25% of the mask size) is determined in the day 228 N mask. Each object in the search region is tested for a match with the template. A match is 229 confirmed if the cross-correlation score exceeds a threshold dynamically scaled (between 0.2 and 230 0.5) based on the object’s size (Fig. 3B-C). The scale of cross-correlation score here is 231 normalized to 0-1. The threshold is defined as the low minimum acceptable cross-correlation 232 score (0.2) plus the difference between the high (0.5) and low (0.2) minimum cross-correlation 233 scores, scaled by 10 times object's size relative to the image. This ensures small objects aren’t 234 penalized for having lower correlation scores and requires larger objects with more prominent 235 features to have higher correlation scores to match. Matched objects are then set to zero in the 236 day N mask to prevent duplicate matching before the process is repeated to find matches for the 237 remaining objects. 238 To address merging objects, labels of all objects in the mask are extracted and a perimeter mask 239 is created by outlining each object’s contours. Neighbors with the most overlapping pixels are 240 identified, and merging is performed if the sizes of both objects fall below 1% of the FOV and their 241 overlapping perimeter exceeds 40%. The smaller object is merged into the larger object. 242 Once matching is complete, the masks are updated with new identification numbers (IDs) (Fig. 243 3B-C). Each pair of matched objects is assigned the same ID. Consolidation of IDs is performed 244 between the updated day N mask from matching with day (N-1) mask and the updated day N 245 mask from matching with day (N+1) mask (Fig. 3C-D). This step links ID pairs that identify the 246 same PDCO in the reference day N mask. 247 Finally, the IDs from the day N-and-day (N+1) matching are propagated to replace old IDs in the 248 day (N-1) mask, according to the ID pairings in the consolidation step (Fig. 3D-E). Masks are 249 updated with new IDs only for matched objects. Unmatched objects are removed from the masks. 250 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint With consistent IDs across day (N-1) and day (N+1) masks, ΔORR for each PDCO can then be 251 calculated by subtracting the normalized ORR of an individual PDCO on day N-1 from the 252 normalized ORR of the same PDCO on day N or (N+1) as in Eq. 5: 253 ∆𝑂𝑅𝑅𝑠𝑖𝑛𝑔𝑙𝑒_𝑃𝐷𝐶𝑂 = 𝑁𝑜𝑟𝑚𝑎𝑙𝑖𝑧𝑒𝑑 𝑂𝑅𝑅𝑃𝐷𝐶𝑂 𝑑𝑎𝑦 𝑁 𝑜𝑟 ሺ𝑁+1ሻ − 𝑁𝑜𝑟𝑚𝑎𝑙𝑖𝑧𝑒𝑑 𝑂𝑅𝑅𝑃𝐷𝐶𝑂 𝑑𝑎𝑦 ሺ𝑁−1ሻ (Eq. 5) 254 Where “Normalized ORRPDCO day N or (N+1)” is the background normalized ORR of individual PDCO 255 on day N or (N+1), and “Normalized ORRPDCO day (N-1)” is the background normalized ORR of the 256 same PDCO measured on day (N-1) prior to treatment, as in Eq. 3. 257 258 Image Background Mask Identification 259 It is inevitable that some PDCOs are not captured with the fine-tuned Cellpose masks. If a simple 260 inverse mask was used, those PDCOs would be included in the background, skewing the 261 accuracy of the background values used for normalization. Therefore, a background algorithm is 262 developed. 263 The algorithm used the following modules and libraries: matplotlib, pandas, tqdm, natsorted, 264 numpy, tifffile, opencv-python, SciPy, scikit-image40–44. The algorithm requires a fine-tuned 265 Cellpose mask, NAD(P)H intensity image, and FAD intensity image to generate a background 266 mask. 267 First, a gradient mask is generated from the intensity images (Fig. 4A) by calculating the 268 normalized gradient magnitude using the Sobel operator and applying Otsu's threshold to 269 separate areas with high spatial intensity gradient, like PDCO edges, from the low-gradient 270

(Fig. 4B). Next, the image is subsampled, keeping every fourth pixel in both the 271 horizontal and vertical directions and these steps are repeated. This is done to increase the 272 intensity gradients and allow detection of more homogenous PDCOs that might have been missed 273 in the first round. While this operation highlights the PDCO edges, the PDCO interiors often have 274 uniform areas of low intensity gradient that are not selected by this operation. Therefore, 275 morphological closing and filling of holes is performed to capture the interior of PDCOs (Fig. 4C). 276 During the development of the algorithm, overmasking occurred frequently at this step. To detect 277 potential overmasking, if the filled gradient mask exceeds 80% of the total image or if the largest 278 object is five times larger than the second largest object, the gradient mask is reverted to the 279 previous mask, before hole filling. The last background mask is then inverted and overlaid with 280 the fine-tuned Cellpose mask (Fig. 4D) to produce a foreground mask (Fig. 4E) to ensure PDCOs 281 of interest are correctly identified as foreground and excluded from the background mask. 282 Next, the top 2% brightest pixels in the image are added to the foreground mask and objects 283 smaller than 25 pixels are removed from the foreground mask, and hole closing and filling 284 operations are performed again. The final background mask is produced by inverting this updated 285 foreground mask. 286 Finally, NAD(P)H and FAD background masks are combined to create the final background mask. 287 If overmasking is detected when both masks are combined, the mask with fewer background 288 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint pixels is selected. If no overmasking is detected, the NAD(P)H and FAD masks are combined 289 using a logical AND operation. 290 In the automated pipeline, the average fluorescence intensity of the background mask pixels in 291 the NAD(P)H image gives the background NAD(P)H value (Ibackground_NAD(P)H) in Eq. 1, and the 292 average fluorescence intensity of the background mask pixels in the FAD image give the 293

FAD value (Ibackground_FAD) in Eq. 2, used to calculate the normalized ORR in Eq. 3. 294 Leading-edge masks 295 Leading-edge masks are only used to calculate final ORR (Eq. 3) and ΔORR_single PDCO (Eq. 296 5) values after tracking and background normalization. In the automated analysis pipeline, a 297 leading-edge mask is created by first eroding a single-PDCO mask by 32 µm, or 20 pixels for 298 Nikon images and 13 pixels for Keyence images due to differences in resolution, followed by 299 subtracting the eroded mask from the original single-PDCO mask. This procedure is the same as 300 the manual pipeline. 301 Statistical Analysis 302 Dice Similarity Coefficient 303 304 The Dice Similarity Coefficient, or Dice Score, was used to calculate the similarity between the 305 fine-tuned Cellpose masks and manual masks. The Dice score is defined as in Eq. 6: 306 𝐷𝑆𝐶 = 2|𝑥∩𝑦| |𝑥|+|𝑦| (Eq. 6) 307 Where |𝑥|is the number of pixels in the first mask, and |𝑦| is the number of pixels in the second 308 mask. |𝑥 ∩ 𝑦|is the number of pixels shared by both masks. The Dice score ranges from 0 to 1, 309 where 0 indicates no overlap between two masks, and 1 indicates complete overlap. 310 Recall 311 Recall was used to assess the accuracy of the tracking algorithm. Recall is defined as in Eq. 7: 312 𝑅𝑒𝑐𝑎𝑙𝑙= 𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠 𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠+𝐹𝑎𝑙𝑠𝑒 𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒𝑠 (Eq. 7) 313 Where true positives are the shared tracks between manual tracking and automated tracking and 314 false negatives are tracks unique to manual tracking. The recall metric measures how many of 315 the manual tracks are correctly detected by the tracking algorithm. 316 Jaccard index 317 Jaccard index was also used to assess the accuracy of the tracking algorithm. Jaccard index is 318 defined as in Eq. 8: 319 𝐽𝑎𝑐𝑐𝑎𝑟𝑑 𝑖𝑛𝑑𝑒𝑥= 𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠 𝐹𝑎𝑙𝑠𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠+𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠+𝐹𝑎𝑙𝑠𝑒 𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒𝑠 (Eq. 8) 320 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint Where true positives are the shared tracks between manual tracking and automated tracking, 321 false positives are tracks unique to automated tracking, and false negatives are tracks unique to 322 manual tracking. The Jaccard index metric measures how many of the unique tracks found in 323 either manual or automated tracking are correctly detected by the tracking algorithm. 324 Median Glass’s Delta (mG𝛥) 325 mGΔ was used to measure the effect size of romidepsin treatment because it provides a more 326 conservative assessment of treatment effects in large sample sizes compared to ANOVA. 327 mGΔ is defined as in Eq. 9 and 10: 328 𝑚𝐺𝛥 = ሺ𝜂𝑡𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡−𝜂𝑐𝑜𝑛𝑡𝑟𝑜𝑙ሻ 𝜎𝑐𝑜𝑛𝑡𝑟𝑜𝑙 (Eq. 9) 329 𝑚𝐺𝛥 = ሺ𝜂𝑑𝑎𝑦ሺ𝑁 𝑜𝑟 𝑁+1ሻ−𝜂𝑑𝑎𝑦ሺ𝑁−1ሻ 𝜎𝑑𝑎𝑦ሺ𝑁−1ሻ (Eq. 10) 330 Where ηtreatment is the median ΔORR (Eq. 4 or 5) of the treatment group, ηcontrol is the median 331 ΔORR of the control group, and σcontrol is the standard deviation of the ΔORR of the control group. 332 In manual analysis, ΔORRpopulation (Eq. 4) was used. In automated analysis, ΔORRsingle_PDCO (Eq. 333 5) was used. 334 For Eq. 10, where ηday (N-1) is the median of ORR (Eq. 3) of the day 0 group, ηday N is the median 335 of ORR (Eq. 3) of the day 1 group, and ηday (N+1) is the median of ORR (Eq. 3) of the day 2 group. 336 σday (N-1) is the standard deviation of the ORR of the day 0 group. 337 Percent Difference 338 Percent difference in ORR between the manual and automated methods was calculated. Percent 339 difference is defined as in Eq. 11: 340 𝑃𝑒𝑟𝑐𝑒𝑛𝑡 𝐷𝑖𝑓𝑓𝑒𝑟𝑒𝑛𝑐𝑒= ሺ𝑂𝑅𝑅𝑎𝑢𝑡𝑜−𝑂𝑅𝑅𝑚𝑎𝑛𝑢𝑎𝑙ሻ 𝑂𝑅𝑅𝑚𝑎𝑛𝑢𝑎𝑙 (Eq. 11) 341 Where ORRauto is the ORR obtained from the automated pipeline, and ORRmanual is the same ORR 342 obtained from the manual pipeline. 343

344 345 Similarity of fine-tuned Cellpose masks and manual masks 346 In prior work, PDCOs were segmented manually using NAD(P)H intensity images in CellProfiler. 347 Manual segmentation is time-consuming and prone to human error, especially when analyzing 348 images on a large scale. Therefore, we adopted Cellpose, a deep-learning based, generalist 349 algorithm, as an automated segmentation tool38. However, existing Cellpose models performed 350 poorly with PDCO segmentation due to their complex, heterogeneous morphology that lacks 351 consistent shapes and features for easy identification36, 37. In addition, wide-field autofluorescence 352 images have low SNR, making segmentation challenging45. To address these issues, we fine-353 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint tuned the cyto3 model using 55 NAD(P)H intensity images and their corresponding masks created 354 manually39. Fifteen NAD(P)H intensity images and their manual masks were tested, including 3 355 images of 30 nM romidepsin treatment taken with the Nikon microscope. A batch processing script 356 using this fine-tuned model then generates PDCO masks, and PDCOs at image borders are 357 removed afterwards. Treatment with romidepsin was chosen for this study to evaluate optical 358 redox imaging’s ability to assess new CRC therapies. Romidepsin is a histone deacetylase 359 (HDAC) inhibitor FDA-approved for cutaneous T-cell lymphoma, peripheral T-cell lymphoma, and 360 multiple myeloma.46–48 It has previously been shown to inhibit tumor growth in CRC tumor 361 models49,50 362 Manual and fine-tuned Cellpose masks demonstrated strong agreement when overlaid for 363 comparison (Fig. 5A, 5B). In the few instances automated and manual masks disagree, PDCOs 364 are out of focus or located at the image borders. The fine-tuned model often includes out-of-focus 365 PDCOs as part of the mask, whereas a trained individual excludes them. Interestingly, despite 366 the training dataset excluding PDCOs at the borders, the fine-tuned model consistently 367 incorporates all PDCOs at the edges of the image. We therefore removed these PDCOs at the 368 image edges using a post-processing script. 369 370 To quantitatively validate the fine-tuned Cellpose masks, a Dice Similarity Coefficient, or dice 371 score, is used to measure the similarity between the manual and automated masks (Fig. 5C, 5D). 372 To demonstrate the generalizability of our approach, images of the same PDCO dishes were 373 acquired on the same day with both Nikon and Keyence microscopes, which have different field-374 of-view dimensions and image resolutions. The Keyence microscope performs automatic stitching 375 of fields of views and therefore captures a larger total area than the Nikon microscope. Both 376 imaging systems achieved mean dice scores exceeding 0.8. The mean dice score for the Nikon 377 microscope (0.84) is somewhat lower than that of the Keyence microscope (0.91), perhaps due 378 to fewer PDCOs in the Nikon images (corresponding to its smaller area images). Overall, this 379 indicates the high reproducibility of our fine-tuned model across different imaging systems, 380 making this a suitable segmentation method for our high-throughput and accessible analysis 381 pipeline. 382 383 Tracking algorithm captures single-PDCO redox changes over time 384 Single-PDCO analysis provides an opportunity to study the impact of tumor heterogeneity on drug 385 resistance. Prior studies have shown that single-PDCO tracking also shows higher sensitivity to 386 treatment response compared to pooled analysis, emphasizing the need for reliable methods to 387 track single PDCOs in a high-throughput drug screen4,26. To address this need, the second 388 innovation in this automated analysis pipeline is an algorithm designed to track single PDCO 389 redox changes across different time points. Two metrics (recall and Jaccard index, Eq. 7 and 8, 390 respectively) were used to test the accuracy of the tracking algorithm, and both metrics show 391 >94% accuracy rate (Fig. 6A-B). 392 Color-coded representative Nikon images illustrate how unique IDs are assigned to each PDCO 393 consistently in images acquired on different days (Fig. 7A). To demonstrate and validate single-394 PDCO tracking with redox values, we first manually segmented, tracked, and calculated the 395 normalized ORR (Eq. 3) of each PDCO on days 0, 1, and 2 of romidepsin or control treatment 396 (Fig. 7B). Only Keyence images were used in this part of the study. For manual tracking, unique 397 IDs were assigned to PDCOs in each image according to the order in which they were segmented, 398 which means that IDs of the same PDCO are not consistent across images acquired on different 399 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint days. Then, a manual process of matching redox ratio values of PDCOs across days was 400 performed. Next, we automatically segmented, tracked, and calculated the normalized ORR (Eq. 401 3, Fig. 7B). We then calculated the percent difference between the ORR obtained from the 402 manual pipeline and the same ORR obtained from the automated pipeline for each individual 403 PDCO (Eq. 11). The percent difference in ORR between the two methods is <25% for all treatment 404 groups (Fig. 7C). mGΔ for all groups are also similar between two pipelines. This suggests there 405 is high agreement between the automated and manual pipelines. 406 Combination of automated segmentation, background normalization, and PDCO 407 tracking accounts for PDCO variability in response 408 409 Next, we compared the performance of the entire automated pipeline against our entire manual 410 pipeline in analyzing ORR changes (Fig. 8). The manual pipeline consists of manual 411 segmentation of PDCOs and sampled background normalization of the ORR with no single-PDCO 412 tracking. The novel automated pipeline includes automated segmentation using fine-tuned 413 Cellpose masks, automated background normalization ORR from NAD(P)H and FAD intensity 414 images, and single-PDCO tracking over the treatment time-course. 415 Here, redox ratio change (ΔORR) is defined differently for the manual and automated pipelines. 416 For the manual analysis ΔORR is calculated using (Eq. 4, ΔORRpopulation) by subtracting the 417 average ORR of a PDCO population on day 0 from the ORR of each individual PDCO on days 1 418 and 2 of treatment. This is due to a lack of tracking capabilities in the manual pipeline. The manual 419 analysis across the Keyence and Nikon systems are consistent, as the day 2, 100 nM romidepsin 420 condition shows the greatest decrease with both microscope systems (Fig. 8A, 8C). 421 We next assessed the performance of the automated pipeline in measuring ΔORR using (Eq. 5, 422 ΔORRsingle_PDCO) where PDCO day 0 ORR is subtracted from day 1 or 2 ORR for each individual 423 PDCO. This is due to the ability to track individual PDCOs over time with the automated pipeline. 424 Notably, almost all standard deviations decreased with single-PDCO tracking (Eq. 5, 425 ΔORRsingle_PDCO) compared to population-level PDCO changes over time (Eq. 4, ΔORRpopulation) 426 (Supplemental Table 3). This suggests that single-PDCO tracking more accurately accounts for 427 PDCO variability compared to population-level calculations, thus increasing the statistical power 428 of drug response measurements, consistent with prior findings26. Comparing across analysis 429 methods, for images acquired on the Nikon, mGΔ increased in all groups with the automated 430 single-PDCO tracking indicating increased sensitivity of the method (Fig. 8A, 8B). For images 431 acquired on the Keyence, mGΔ increased in all groups with the automated single-PDCO tracking 432 except day 2, 100 nM romidepsin condition (Fig. 8C, 8D). Importantly, both microscopes and 433 analysis methods show that the day 2, 100 nM romidepsin group has the largest decrease in 434 ΔORR. Furthermore, the automated algorithm also removes human bias and increases 435 throughput, thus improving the accessibility of wide-field optical redox imaging for testing PDCO 436 drug response. 437 438 Automated Pipeline Significantly Reduces Analysis Time Compared to Manual Pipeline 439 We next compared the time required to complete analysis of 9 NAD(P)H and FAD intensity image 440 pairs with the manual pipeline to that of the automated pipeline (Table 1). These images were 441 acquired with the Nikon microscope only. Due to a different FOV size, Keyence images would 442 likely take ~4 times longer per image for manual analysis. Manual segmentation time is defined 443 as the time between the opening of an image to the time after all PDCOs are segmented. Manual 444

time is defined as the time between the opening of an image and the time after 5 445 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint PDCO-free regions are selected. Manual tracking time was defined as the time required to match 446 IDs for every PDCO across the three NAD(P)H images within a time series, and record the IDs in 447 an Excel sheet, repeated over 3 time-series. In the case of manual tracking, two experts 448 completed the entire dataset independently to ensure accuracy, which required ~2 hours per 449 three-image time series x 3 time-series (6 hours) per person. The total manual pipeline time 450 includes manual segmentation, manual background, and manual tracking time. 451 452 The time required to complete analysis of the same 3 sets of 3 time-series images of both 453 NAD(P)H and FAD (9 images for NAD(P)H and 9 images for FAD) with the automated image 454 analysis pipeline was recorded using the Python DateTime module (Table 1). Automated 455 segmentation time was defined as the execution time of the fine-tuned Cellpose model batch 456 processing script. Automated tracking time is defined as the time required to match IDs for every 457 PDCO across the three images (Fig. 2). Automated background time is defined as the execution 458 time of the script that includes the background identification algorithm (Fig. 3). The total 459 automated pipeline time includes automated segmentation, automated background, and 460 automated tracking time. 461 462 Segmentation Time

Time Tracking Time Total Pipeline Time Manual 1 hour 9 min > 6 hours > 7 hours Automated 37 s 2.3 min 25 s 3.3 min Table 1: The time required to complete analysis of 3 sets of 3 time-series images of both NAD(P)H and FAD (9 463 images for NAD(P)H and 9 images for FAD) with either the manual image analysis pipeline or automated image 464 analysis pipeline. PDCOs were tracked over 3 time points, and images were collected with the Nikon system. 465 466

467 Wide-field optical redox imaging is a nondestructive approach to assess patient treatment 468 response by tracking longitudinal metabolic changes in PDCOs at the single-PDCO level. This 469 presents an opportunity to apply wide-field optical redox imaging to predict the most effective 470 therapies and identify new therapeutic candidates. To facilitate the expansion of wide-field optical 471 redox imaging as a functional drug screening method in PDCOs, a more accessible and high-472 throughput image analysis pipeline is needed to meet the demands of large-scale drug screens. 473 PDCO image analysis is a challenging computational problem due to the variability in PDCO 474 shape, size, and position across focal planes. Autofluorescence images have low SNR, making 475 analysis even more challenging45. Several open-source programs, such as OrganoID, MOrgAna, 476 OrgaExtractor, and deepOrganoid, have been developed to address PDCO segmentation and 477 tracking28–31. However, these tools are not designed for autofluorescence images, making them 478 unsuitable for assessing wide-field redox imaging data sets. 479 In some instances, the automated image analysis pipeline developed here disagrees with the 480 manual image analysis pipeline. When comparing between the two pipelines, mGΔ values are 481 similar but not the same. Automated background identification covers a larger area than sampled 482

identification and likely provides a more accurate representation of the background 483 than only 5 PDCO-free regions selected manually. 484 Our automated pipeline not only achieves similar sensitivity to drug response in PDCOs as our 485 manual pipeline but also allows us to resolve single-PDCO treatment responses at a significantly 486 faster processing time (>127 times faster, Table 1), saving time and limiting user errors. This 487 automated pipeline requires no computational expertise, thereby providing an accessible PDCO 488 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint image analysis tool. Furthermore, automated single-PDCO tracking enables paired statistical 489 tests that are more sensitive than unpaired population level tests. However, the automated 490 pipeline struggles in some cases. First, because our pipeline relies on dynamic thresholds that 491 account for both position and morphology of PDCOs to track over time, it is less reliable when a 492 PDCO moves or changes in morphology beyond pre-determined thresholds. Second, although 493 the algorithms were designed specifically for autofluorescence images, they still struggle where 494 signals are too low. To improve the generalizability and accuracy of the pipeline, existing deep 495 learning methods may help address these limitations in image variation28–31. Third, the automated 496 pipeline is generalizable across the two imaging systems tested (Nikon and Keyence), but image 497 acquisition parameters such as the objective lens used may affect efficacy. Finally, the middle 498 day image (day N), compared to the control image (day N-1), was chosen as the reference for 499 PDCO tracking to minimize the gap between neighboring days on both sides, which increases 500 the probability of matching across all three days. However, if PDCOs are missing due to treatment 501 effects, this choice of reference day could generate errors and therefore the imaging frequency 502 should be increased in those cases. 503 Overall, this study presents an improved image analysis pipeline that automates PDCO 504 segmentation, single-PDCO tracking, and background correction for wide-field redox imaging. 505 This improves the throughput and robustness of PDCO image analysis and represents a step 506 towards our goal to enhance the accessibility of wide-field optical redox imaging for PDCO drug 507 screening. 508

509 510 1. Cancer of the Colon and Rectum - Cancer Stat Facts. SEER 511 https://seer.cancer.gov/statfacts/html/colorect.html. 512 2. Saoudi González, N. et al. Unravelling the Complexity of Colorectal Cancer: Heterogeneity, 513 Clonal Evolution, and Clinical Implications. Cancers (Basel) 15, 4020 (2023). 514 3. Demicco, M., Liu, X.-Z., Leithner, K. & Fendt, S.-M. Metabolic heterogeneity in cancer. Nat 515 Metab 6, 18–38 (2024). 516 4. Skala, M. C., Deming, D. A. & Kratz, J. D. Technologies to Assess Drug Response and 517 Heterogeneity in Patient-Derived Cancer Organoids. Annual Review of Biomedical 518 Engineering 24, 157–177 (2022). 519 5. Kopper, O. et al. An organoid platform for ovarian cancer captures intra- and interpatient 520 heterogeneity. Nat Med 25, 838–849 (2019). 521 6. Vlachogiannis, G. et al. Patient-derived organoids model treatment response of metastatic 522 gastrointestinal cancers. Science 359, 920–926 (2018). 523 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint 7. Quinn, K. P. et al. Quantitative metabolic imaging using endogenous fluorescence to detect 524 stem cell differentiation. Sci Rep 3, 3432 (2013). 525 8. Georgakoudi, I. & Quinn, K. P. Optical imaging using endogenous contrast to assess 526 metabolic state. Annu Rev Biomed Eng 14, 351–367 (2012). 527 9. Cannon, T. M., Shah, A. T., Walsh, A. J. & Skala, M. C. High-throughput measurements of 528 the optical redox ratio using a commercial microplate reader. J Biomed Opt 20, 010503 529 (2015). 530 10. Walsh, A. J. et al. Quantitative Optical Imaging of Primary Tumor Organoid Metabolism 531 Predicts Drug Response in Breast Cancer. Cancer Research 74, 5184–5194 (2014). 532 11. Pauli, C. et al. Personalized In Vitro and In Vivo Cancer Models to Guide Precision 533 Medicine. Cancer Discov 7, 462–477 (2017). 534 12. Ganesh, K. et al. A rectal cancer organoid platform to study individual responses to 535 chemoradiation. Nat Med 25, 1607–1614 (2019). 536 13. Driehuis, E., Kretzschmar, K. & Clevers, H. Establishment of patient-derived cancer 537 organoids for drug-screening applications. Nat Protoc 15, 3380–3409 (2020). 538 14. Ooft, S. N. et al. Patient-derived organoids can predict response to chemotherapy in 539 metastatic colorectal cancer patients. Sci Transl Med 11, eaay2574 (2019). 540 15. Tveit, K. M. & Pihl, A. Do cell lines in vitro reflect the properties of the tumours of origin? A 541 study of lines derived from human melanoma xenografts. Br J Cancer 44, 775–786 (1981). 542 16. Esquenet, M., Swinnen, J. V., Heyns, W. & Verhoeven, G. LNCaP prostatic 543 adenocarcinoma cells derived from low and high passage numbers display divergent 544 responses not only to androgens but also to retinoids. J Steroid Biochem Mol Biol 62, 391–545 399 (1997). 546 17. Daniel, V. C. et al. A primary xenograft model of small-cell lung cancer reveals irreversible 547 changes in gene expression imposed by culture in vitro. Cancer Res 69, 3364–3373 (2009). 548 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint 18. Sambuy, Y. et al. The Caco-2 cell line as a model of the intestinal barrier: influence of cell 549 and culture-related factors on Caco-2 cell functional characteristics. Cell Biol Toxicol 21, 1–550 26 (2005). 551 19. Wilding, J. L. & Bodmer, W. F. Cancer cell lines for drug discovery and development. 552 Cancer Res 74, 2377–2384 (2014). 553 20. Phan, N. et al. A simple high-throughput approach identifies actionable drug sensitivities in 554 patient-derived tumor organoids. Commun Biol 2, 1–11 (2019). 555 21. Lakowicz, J. R., Szmacinski, H., Nowaczyk, K. & Johnson, M. L. Fluorescence lifetime 556 imaging of free and protein-bound NADH. Proc Natl Acad Sci U S A 89, 1271–1275 (1992). 557 22. Roh, T. T. et al. Predicting DNA damage response in non-small cell lung cancer organoids 558 via simultaneous label-free autofluorescence multiharmonic microscopy. Redox Biology 75, 559 103280 (2024). 560 23. Xu, H. N., Tchou, J., Feng, M., Zhao, H. & Li, L. Z. Optical redox imaging indices 561 discriminate human breast cancer from normal tissues. J. Biomedical Optics 21, 114003 562 (2016). 563 24. XU, H. N., WU, B., NIOKA, S., CHANCE, B. & LI, L. Z. QUANTITATIVE REDOX 564 SCANNING OF TISSUE SAMPLES USING A CALIBRATION PROCEDURE. J Innov Opt 565 Health Sci 2, 375–385 (2009). 566 25. Ranji, M. et al. Fluorescence spectroscopy and imaging of myocardial apoptosis. J Biomed 567 Opt 11, 064036 (2006). 568 26. Gil, D. A., Deming, D. & Skala, M. C. Patient-derived cancer organoid tracking with wide-569 field one-photon redox imaging to assess treatment response. J Biomed Opt 26, 036005 570 (2021). 571 27. Gillette, A. et al. Wide-Field Optical Redox Imaging with Leading-Edge Detection Enables 572 Assessment of Treatment Response and Heterogeneity in Patient-Derived Cancer 573 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint Organoids. Cancer Res https://doi.org/10.1158/0008-5472.CAN-24-4958 (2025) 574 doi:10.1158/0008-5472.CAN-24-4958. 575 28. Matthews, J. M. et al. OrganoID: A versatile deep learning platform for tracking and analysis 576 of single-organoid dynamics. PLOS Computational Biology 18, e1010584 (2022). 577 29. Gritti, N. et al. MOrgAna: accessible quantitative analysis of organoids with machine 578 learning. Development 148, dev199611 (2021). 579 30. Park, T. et al. Development of a deep learning based image processing tool for enhanced 580 organoid analysis. Sci Rep 13, 19841 (2023). 581 31. Powell, R. T. et al. deepOrganoid: A brightfield cell viability model for screening matrix-582 embedded organoids. SLAS Discovery 27, 175–184 (2022). 583 32. Pasch, C. A. et al. Patient-derived cancer organoid cultures to predict sensitivity to 584 chemotherapy and radiation. Clin Cancer Res 25, 5376–5387 (2019). 585 33. Kratz, J. D. et al. Integrating Subclonal Response Heterogeneity to Define Cancer Organoid 586 Therapeutic Sensitivity. Preprint at https://doi.org/10.1101/2021.10.15.464556 (2021). 587 34. Jagiella, N., Müller, B., Müller, M., Vignon-Clementel, I. E. & Drasdo, D. Inferring Growth 588 Control Mechanisms in Growing Multi-cellular Spheroids of NSCLC Cells from Spatial-589 Temporal Image Data. PLoS Comput Biol 12, e1004412 (2016). 590 35. Cui, X., Hartanto, Y. & Zhang, H. Advances in multicellular spheroids formation. J R Soc 591 Interface 14, 20160877 (2017). 592 36. Gillette, A. et al. Wide-field optical redox imaging with leading-edge detection for 593 assessment of patient-derived cancer organoids. bioRxiv 2024.12.23.630148 (2024) 594 doi:10.1101/2024.12.23.630148. 595 37. Theodossiou, A., Martinez, J. & Walsh, A. J. Fast autofluorescence imaging to evaluate 596 dynamic changes in cell metabolism. J Biomed Opt 29, 126501 (2024). 597 38. Pachitariu, M. & Stringer, C. Cellpose 2.0: how to train your own model. Nat Methods 19, 598 1634–1641 (2022). 599 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint 39. Stringer, C. & Pachitariu, M. Cellpose3: one-click image restoration for improved cellular 600 segmentation. Nat Methods 22, 592–599 (2025). 601 40. Hunter, J. D. Matplotlib: A 2D Graphics Environment. Computing in Science & Engineering 602 9, 90–95 (2007). 603 41. Virtanen, P. et al. SciPy 1.0: fundamental algorithms for scientific computing in Python. Nat 604

17, 261–272 (2020). 605 42. Walt, S. van der et al. scikit-image: image processing in Python. PeerJ 2, e453 (2014). 606 43. Harris, C. R. et al. Array programming with NumPy. Nature 585, 357–362 (2020). 607 44. McKinney, W. Data Structures for Statistical Computing in Python. in 56–61 (Austin, Texas, 608 2010). doi:10.25080/Majora-92bf1922-00a. 609 45. Riendeau, J. M. et al. Cellpose as a reliable method for single-cell segmentation of 610 autofluorescence microscopy images. Sci Rep 15, 5548 (2025). 611 46. Whitehead, R. P. et al. Phase II trial of romidepsin (NSC-630176) in previously treated 612 colorectal cancer patients with advanced disease: a Southwest Oncology Group study 613 (S0336). Invest New Drugs 27, 469–475 (2009). 614 47. Woo, S. et al. Population pharmacokinetics of romidepsin in patients with cutaneous T-cell 615 lymphoma and relapsed peripheral T-cell lymphoma. Clin Cancer Res 15, 1496–1503 616 (2009). 617 48. Pu, J. et al. Exploring the role of histone deacetylase and histone deacetylase inhibitors in 618 the context of multiple myeloma: mechanisms, therapeutic implications, and future 619 perspectives. Experimental Hematology & Oncology 13, 45 (2024). 620 49. Shi, Y. et al. Romidepsin (FK228) regulates the expression of the immune checkpoint ligand 621 PD-L1 and suppresses cellular immune functions in colon cancer. Cancer Immunol 622 Immunother 70, 61–73 (2021). 623 50. Li, J., Pan, J., Wang, L., Ji, G. & Dang, Y. Colorectal Cancer: Pathogenesis and Targeted 624 Therapy. MedComm 6, e70127 (2025). 625 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint Author Contributions 626 AH: design, analysis, and interpretation of data, creation of new software used in the 627 work, and drafted the work; KS: design, analysis, and interpretation of data, creation of 628 new software used in the work, and substantively revised the work; AAG: conception, 629 design, analysis, and interpretation of data; SU: acquisition of data; AES: acquisition of 630 data, WZ: analysis of data and creation of new software used in the work; DD: 631 conception and design of the work; MCS: conception, design, and interpretation of data, 632 and substantively revised the work. All authors have read and approved this manuscript. 633 634

635 We thank Skala and Deming lab members for their helpful feedback on the figures and 636 manuscript. We thank Matthew Stefely for his contributions to figure editing. We also thank Dr. 637 Lang Wang for his contributions to the early development of our background algorithm. 638 639 Data/Code availability 640 Dataset used in this study is available upon reasonable request. The code presented in this 641 study can be found here: 642 https://github.com/skalalab/Auto_WF_PDCO_Pipeline 643 644 Additional Information (Disclosures and Funding Sources) 645 MCS is an advisor for Elephas Biosciences. This entity had no input in the study design, 646 analysis, manuscript preparation or decision to submit for publication. 647 648 The author(s) declare that financial support was received for the research and/or publication of 649 this article. This research was supported by the National Institutes of Health grants R01 650 CA278051, R01 CA272855, R37 CA226526, Morgridge Institute for Research, Carol Skornicka 651 Chair of Biomedical Imaging. Additionally, the Deming laboratory is funded by the ACI/Schwenn 652 Family Professorship, JD Fluno Family Colorectal Cancer Precision Medicine Program, and the 653 V Foundation. 654 655 656 657 658 659 660 661 662 663 664 665 666 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint 667 Figure 1: Representative NAD(P)H intensity image of sampled background normalization from the manual pipeline. An 668 individual selects five PDCO-free regions (white) and uses the average fluorescence intensity as the background value 669 to normalize NAD(P)H and FAD intensity images. Color scale bar reflects intensity scaled 0-255 for visualization. 670 Numbered PDCO-free regions are illustrated in white for visualization only and have intensity values near zero. 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint 700 701 702 Figure 2: Workflow of the automated wide-field PDCO ORR image analysis pipeline. (A) The fine-tuned Cellpose 703 model was used to generate new masks with a batch-processing script. A separate script is used to remove incomplete 704 PDCO masks with pixels touching the borders of the ROI (not shown). (B) Day (N-1), day N, and day (N+1) masks can 705 be used as inputs for the tracking algorithm, which assigns matching objects in each mask with the same unique ID 706 based on cross-correlation scores and location. (C) Masks with matched unique IDs, with NAD(P)H and FAD intensity 707 images, can then be used as inputs for the background pixel identification algorithm. (D) Leading-edge masks are 708 created. (E) NAD(P)H and FAD intensity images are background normalized and used to generate the ORR image. (F) 709 The ORR mean values are calculated in the leading-edge mask for each PDCO across timepoints and are saved to 710 the output CSV file. 711 712 713 714 715 716 717 718 719 720 721 722 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint 723 Figure 3: General workflow for the algorithm to track PDCOs within NAD(P)H images over multiple time-points. 724 Coarse image registration occurs before the following steps (not shown). (A-B) Light blue arrows indicate the removal 725 of small objects likely to be debris. (B-D) Red arrows indicate the object matching and ID consolidation step where the 726 old IDs (numbers) are updated to new IDs (letters). (B-C) Red circled numbers indicate object matching based on cross-727 correlation, first between day (N-1) and day N, then day (N+1) and day N. (C-D) ID consolidation occurs between the 728 two updated day N masks, linking ID pairs (e.g., a d, b e, c f). (C-E) Purple arrows indicate paired IDs are 729 propagated to ensure updated IDs are consistent across all days for each PDCO. Numbered and lettered shapes 730 represent PDCOs. Numbers and letters represent IDs, or unique identification labels. Numbered IDs are unique to each 731 PDCO in each image. Letter IDs are unique to each tracked PDCO across two or more time points. Changes in shape 732 size and orientation from day (N-1) to day (N+1) illustrate changes in PDCOs after treatment. 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint 756 Figure 4: General workflow for the algorithm to automatically detect background pixels within 757 autofluorescence images. (A) Representative input NAD(P)H intensity image. NAD(P)H and FAD images are 758 separately processed to identify background pixels for each image. Color scale bar scaled to 0-255. (B) Gradient mask 759 generated after Sobel operation. (C) Mask after applying closing-and-filling-holes operation. (D) Fine-tuned Cellpose 760 mask (E) Final background mask generated after subtracting the area covered by the fine-tuned Cellpose mask. 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint 784 785 Figure 5. Dice score measurements reveal high agreement between manually segmented masks and Cellpose 786 fine-tuned masks. Representative images of wide-field NAD(P)H intensity images of CRC PDCOs before (day 0) and 787 2 days after 30 nM romidepsin treatment (day 2) for (A) Nikon microscope and (B) Keyence microscope images. Masks 788 were overlaid for manually segmented (magenta) and Cellpose fine-tuned masks (cyan) where more overlap (white) 789

in a higher dice score. The Dice Similarity Coefficient, or Dice Score, is computed by taking twice the white pixel 790 area and dividing it by the combined pixel areas of the magenta and cyan masks (Eq. 6). Nikon scale bar = 400 µm. 791 Keyence scale bar = 2000 µm. The same wells of PDCOs were imaged with both the Keyence and Nikon systems, 792 however, on the Keyence system automatic stitching was performed with a 30% overlap between adjacent FOV, 793 yielding one large FOV per well. For (C) Nikon Dice Score, each dot represents a FOV, 4 acquired per PDCO well (n 794 = 36). For (D) Keyence Dice Score, each dot represents a full-field image of the same PDCO well (n = 9). Center line 795 is mean. Error bars are standard deviation. 796 797 798 799 800 801 802 803 804 805 806 807 808 809 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint 810 Figure 6: Both recall and Jaccard index show high accuracy for the tracking algorithm compared to manual 811 tracking. 3-day imaging time course of CRC PDCOs treated with (A) control media or (B) 100nM romidepsin acquired 812 with the Keyence microscope. White represents tracks shared by both manual tracking and automated tracking. 813 Magenta represents tracks unique to manual tracking. Cyan represents tracks unique to automated tracking. For 814 Control: Shared tracks n = 69. Unique manual tracks n = 0. Unique automated tracks n = 0. For 100 nM romidepsin: 815 Shared tracks n = 68. Unique manual tracks n = 2. Unique automated tracks n = 2. Scale bar = 2000 µm. 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint 835 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint Figure 7. The automated pipeline shows strong agreement with the manual pipeline in tracking single PDCO 836 wide-field redox changes. (A) Representative Nikon images of PDCO-matched masks generated from the tracking 837 algorithm. PDCOs matched across days are highlighted in the same color. White arrows indicate an example of where 838 PDCOs have shifted over time. (B) Keyence ORRs (Eq. 3) are plotted for single PDCOs segmented, tracked, and 839 calculated manually or automatically. mGΔ is used to measure effect size of treatment. (C) Percent differences between 840 manually and automatically calculated ORR are plotted. PDCO numbers for control n = 69; 30 nM romidepsin n = 79; 841 100 nM romidepsin n = 70. 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint 888 889 Figure 8. ORR changes obtained with manual and automated pipelines. (A,C) ΔORRpopulation (Eq. 4) were 890 measured using the manual pipeline, which includes manual segmentation and sampled background normalization as 891 previously described, but with no single PDCO tracking, in a (A) Nikon or (C) Keyence microscope. In the manual 892 pipeline, ΔORRpopulation (Eq. 4) was calculated by subtracting the average ORR of a PDCO population before romidepsin 893 (Romi) treatment on day 0 (population day 0) from the ORR of each PDCO in the same population after romidepsin 894 treatment (Romi) on day 1 or 2 as specified. (B,D) ΔORRsingle_PDCO (Eq. 5) were measured using the entire automated 895 pipeline including segmentation, background normalization, and single PDCO tracking, in a (B) Nikon or (D) Keyence 896 microscope. Here, the ORR of each day 0 PDCO (matched day 0 PDCO) is subtracted from day 1 or day 2 ORR of the 897 same PDCO, excluding unmatched organoids. Number of PDCOs used in (A) and (C) are described in Supplemental 898 Table 1. Supplemental Table 2 describes the number of PDCOs included in single-PDCO tracking analysis (B, D). mGΔ 899 (Eq. 10) is used to measure effect size of treatment. Center line is mean. Error bars are standard deviation. 900 901 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted October 31, 2025. ; https://doi.org/10.1101/2025.10.29.685427doi: bioRxiv preprint