Colon: Anatomy, Imaging, and Molecular Insights

• Colon is a 1.5–1.8 m long part of the large intestine characterized by complex folds and patient-specific deformations essential in diagnostic and navigational procedures.
• Advanced imaging and computational methods, including CT segmentation, CNN-based histopathology, and 3D reconstruction, enhance clinical navigation and diagnostic accuracy.
• Molecular studies and algebraic modeling of colon ideals offer diverse insights into colon cancer pathways and innovative mathematical analogs.

The colon is a 1.5–1.8 m long tubular organ of the large intestine extending from the cecum through the ascending, transverse, descending, and sigmoid segments to the rectum and anus; its wall contains haustral folds, and its lumen can adopt highly tortuous, patient-specific configurations (Oda et al., 2018). In current research, the term spans several technically distinct domains: anatomical modeling for colonoscopy navigation, histopathological image analysis, CT-based segmentation, monocular and synthetic 3D reconstruction, multimodal lesion annotation in full-procedure videos, molecular studies of colon cancer, and, in commutative algebra, the ideal quotient known as a colon ideal (Oda et al., 2018, Booth et al., 12 Mar 2026).

1. Anatomy, deformation, and navigation during colonoscopy

During colonoscopy, the endoscopist must navigate sharp bends and narrow passages formed by the haustral folds. Accurate, up-to-date knowledge of the colon’s three-dimensional shape is useful for avoiding looping, minimizing blind advancement into sharp turns that risk wall perforation, and guiding the physician to previously identified lesions or polyps (Oda et al., 2018). A central complication is that the colon deforms during insertion: as the colonoscope is inserted and torqued, it pushes against the colonic wall and mesentery, causing local stretches, bends, and sometimes global shape changes. The amount and direction of deformation depend on insertion speed, applied torque, patient posture, and anatomical variation, so preoperative CT colonography does not directly represent the intra-procedural geometry (Oda et al., 2018).

A formalization used in colonoscope tracking denotes the undeformed colon centerline or marker set by $x \in \Omega$ and the deformed colon by $y' \in \mathbb{R}^3$, with deformation field $D:\Omega \to \mathbb{R}^3$ satisfying

$y' = x + D(x).$

The estimation problem is then to learn a mapping from the measured colonoscope shape ${\bf X}=({\bf x}_1,\dots,{\bf x}_N)$ to $D(\cdot)$ (Oda et al., 2018). In the reported regression-forest framework, the predicted deformation is represented as an ensemble of decision trees, with input features including local curvature or bending computed from nearby electromagnetic sensing points, tangent orientations, distances to the nearest measured points, and global insertion depth (Oda et al., 2018).

In the phantom study, CT colonography provided the undeformed colon model, a Kinect v2 system recorded 12 surface markers on a colon phantom as ground truth, and an electromagnetic sensor with six sensing points captured the colonoscope shape. Six insertions trained 12 regressors with 100 trees each, and a seventh insertion was held out for testing. The mean deformation error was on the order of a few millimeters, approximately 3–5 mm, and the estimated marker positions lay on or near the phantom surface across representative insertion states, although a few outliers outside the surface indicated the need for biomechanical constraints (Oda et al., 2018). This establishes deformation estimation as a geometric prerequisite for safer real-time navigation rather than as a purely post hoc registration problem.

2. Histopathology, datasets, and computational pathology

The colon subset of LC25000 is a balanced binary histopathology dataset containing two classes after augmentation: colon_aca for colon adenocarcinoma and colon_n for benign colon tissue, with 5,000 images per class and 10,000 colon images in total (Borkowski et al., 2019). Before augmentation, the subset contained 250 adenocarcinoma and 250 benign images. The source material was de-identified, HIPAA-compliant pathology glass slides, stored as 24-bit RGB JPEG images originally captured at $1024 \times 768$ pixels; the preprocessing extracted a central square region of $768 \times 768$ pixels, and augmentation used Augmentor with random rotation of $\pm 25^\circ$ at probability $1.0$, random horizontal flip at probability $y' \in \mathbb{R}^3$0, and random vertical flip at probability $y' \in \mathbb{R}^3$1 (Borkowski et al., 2019). Ground truth derived from original surgical pathology diagnoses, and all cases were validated by at least one board-certified pathologist, but no formal inter-observer agreement, $y' \in \mathbb{R}^3$2-statistic, second-read, or consensus adjudication was reported (Borkowski et al., 2019).

A colon-specific CNN study based on LC25000 resized all inputs to $y' \in \mathbb{R}^3$3, normalized pixel values to $y' \in \mathbb{R}^3$4, and applied random shear and zoom transformations during training (Mangal et al., 2020). The train/validation/test split was 80/10/10, yielding 8,000 training images, 1,000 validation images, and 1,000 test images, but no patient-level identifiers were available, so evaluation was image-level rather than patient-level (Mangal et al., 2020). The reported architecture was a shallow network with three convolutional layers using “same” padding and ReLU activations, followed by flattening, a 512-unit dense layer with dropout rate $y' \in \mathbb{R}^3$5, and a 2-unit softmax output layer. Optimization used RMSprop with learning rate $y' \in \mathbb{R}^3$6, $y' \in \mathbb{R}^3$7, $y' \in \mathbb{R}^3$8, batch size 32, and 100 epochs (Mangal et al., 2020). The final held-out test accuracy was approximately 96%, with training accuracy $y' \in \mathbb{R}^3$9 and validation accuracy $D:\Omega \to \mathbb{R}^3$0 (Mangal et al., 2020).

A different line of work addresses gland segmentation rather than slide-level classification. On the GlaS@MICCAI’15 dataset, one method first applied Reinhard color normalization to H&E images, then used a hierarchical three-level random-forest cascade on non-overlapping patches of sizes $D:\Omega \to \mathbb{R}^3$1, $D:\Omega \to \mathbb{R}^3$2, and $D:\Omega \to \mathbb{R}^3$3 (Khatun et al., 2019). Features combined RGB histograms, local mean and standard-deviation based 2D joint histograms, and all 14 standard Haralick GLCM descriptors per color channel. Level 1 classified patches into $D:\Omega \to \mathbb{R}^3$4Gland, Non-Gland, Mix$D:\Omega \to \mathbb{R}^3$5, while subsequent levels recursively refined only the “Mix” patches; post-processing used Canny edge detection, edge dilation with an octagonal structuring element, and superposition on the raw label map to smooth gland boundaries (Khatun et al., 2019). The reported overall patch-classification accuracy averaged over five runs was $D:\Omega \to \mathbb{R}^3$6, exceeding the reported random forest, k-NN, and SVM baselines in that study (Khatun et al., 2019).

A recurring misconception is that large augmented histology datasets automatically imply broad clinical generalization. The LC25000 colon subset is explicitly limited by single-center slide origin, geometric-only augmentation, absence of stain normalization, lack of granular labels such as tumor grade or subtype, and the absence of an independent hold-out test set beyond augmented splits (Borkowski et al., 2019).

3. CT segmentation and high-resolution annotation

CT-based segmentation work treats the colon as a high-variability structure whose geometry challenges generic abdominal segmentors. HQColon constructed a high-resolution colon dataset from the TCIA CT Colonography collection by filtering 3,451 scans down to 435 annotated scans (Finocchiaro et al., 28 Feb 2025). Air-filled colon was isolated by thresholding at intensities $D:\Omega \to \mathbb{R}^3$7 HU, selecting a seed by scanning a $D:\Omega \to \mathbb{R}^3$8-pixel strip around the anterior–posterior midpoint on axial slices 50–250, and then performing 26-neighborhood 3D region growing with $D:\Omega \to \mathbb{R}^3$9 HU. Connected components with $y' = x + D(x).$0 or $y' = x + D(x).$1 were discarded, and a rapid expert visual check removed collapsed segments or small-bowel leaks (Finocchiaro et al., 28 Feb 2025).

Fluid-filled colon was then labeled by interactive machine learning in RootPainter. The pipeline dilated the TotalSegmentator colon mask by 35 voxels, exported seven axial slices per scan for corrective annotation, iteratively trained a U-shaped CNN, and stopped when Dice improvements diminished, with roll-mean Dice $y' = x + D(x).$2 after approximately 80 minutes and total annotation time of 215 minutes for 390 slices (Finocchiaro et al., 28 Feb 2025). Post-processing removed islands smaller than 2,000 voxels or farther than 2 mm from the air-filled surface, applied gravity-based filtering, hole filling, Gaussian smoothing, and sagittal bridging. The final dataset contained 435 scans split into 290 training and 145 test cases, with approximately 4 minutes of expert time per scan (Finocchiaro et al., 28 Feb 2025).

The automatic segmentor was a full-resolution 3D nnU-Net v2 with five down-sampling and five up-sampling levels, kernel size $y' = x + D(x).$3, feature maps $y' = x + D(x).$4 at the bottleneck, Dice plus cross-entropy loss, SGD with momentum $y' = x + D(x).$5, initial learning rate $y' = x + D(x).$6, 1,000 epochs, and batch size 2 on an NVIDIA RTX 4090 (Finocchiaro et al., 28 Feb 2025). On 145 test scans, HQColon achieved $y' = x + D(x).$7 mm and $y' = x + D(x).$8 mm, compared with TotalSegmentator’s $y' = x + D(x).$9 mm and ${\bf X}=({\bf x}_1,\dots,{\bf x}_N)$0 mm (Finocchiaro et al., 28 Feb 2025). The study reported that results were virtually identical with raw versus masked input and before versus after tiny-island removal, indicating robustness of the nnU-Net configuration (Finocchiaro et al., 28 Feb 2025).

A related dataset extension of the Medical Segmentation Decathlon added colon masks to CT studies with colorectal cancer labels. Starting from 126 abdominal MSCT studies and 842 additional TotalSegmentator CT volumes, a radiologist with 15 years of abdominal CT experience validated colon boundaries and categorized volumes into “Good,” “Bad – cropped,” and “Excluded” subsets (Chernenkiy et al., 2024). For model development, 100 “good” volumes were used in five-fold cross-validation. Reported Dice scores were ${\bf X}=({\bf x}_1,\dots,{\bf x}_N)$1 for colon and ${\bf X}=({\bf x}_1,\dots,{\bf x}_N)$2 for tumor with PlainUnet, and ${\bf X}=({\bf x}_1,\dots,{\bf x}_N)$3 for colon and ${\bf X}=({\bf x}_1,\dots,{\bf x}_N)$4 for tumor with an nnU-Net low-resolution baseline (Chernenkiy et al., 2024). Including low-quality scans degraded colon segmentation by 3–4 percentage points in Dice, underscoring the sensitivity of colon labeling to incomplete visualization and challenging anatomy (Chernenkiy et al., 2024).

4. 3D reconstruction, synthetic data, and coverage estimation

Coverage estimation addresses how much of the inner colon surface has been visually inspected during colonoscopy. One formulation defines segment coverage as

${\bf X}=({\bf x}_1,\dots,{\bf x}_N)$5

where ${\bf X}=({\bf x}_1,\dots,{\bf x}_N)$6 is the surface area reconstructed from the video and ${\bf X}=({\bf x}_1,\dots,{\bf x}_N)$7 is the surface area of the completed mesh after inferring missing geometry (Muhlethaler et al., 2022). The pipeline first segments the colon into contiguous sections of fixed arc length ${\bf X}=({\bf x}_1,\dots,{\bf x}_N)$8, for example 5–7 cm, by estimating a centerline with a learned 3D U-Net and a minimal-path fast-marching algorithm operating on a ${\bf X}=({\bf x}_1,\dots,{\bf x}_N)$9 volumetric heatmap (Muhlethaler et al., 2022). For each segment, marching cubes extracts a watertight completed mesh, a partial mesh is obtained by thresholding distances to the observed point cloud, and triangle areas are summed to compute $D(\cdot)$0, $D(\cdot)$1, and the missing area (Muhlethaler et al., 2022).

Training used three full-colon meshes hand-segmented from CT colonograms and additional synthetic colonoscopy videos generated inside two synthetic colon meshes. The full meshes were split into segments of length $D(\cdot)$2, yielding 10,200 training segments, 3,000 validation segments, and 1,200 held-out test segments (Muhlethaler et al., 2022). The principal metric was mean absolute coverage error,

$D(\cdot)$3

On the held-out segments, MAE was approximately 0.03 without added noise and approximately 0.06 with Gaussian noise of $D(\cdot)$4 mm (Muhlethaler et al., 2022). On a 3D-printed colon phantom and a real optical colonoscopy video from Colon10K, the method produced qualitative visualizations in which under-inspected regions appeared in red and well-covered segments in green (Muhlethaler et al., 2022).

Synthetic data have become a complementary route for solving the scarcity of dense ground truth in endoscopic 3D reconstruction. RealSynCol generated 28,130 frames from ten patient CT-derived colon geometries rendered in Blender 3.6.9 with a pinhole camera model, a proximal xenon-style light source, vascular textures synthesized from the SUN clinical image database, and physically based rendering for specular highlights (Lena et al., 9 Feb 2026). Each of 20 sequences was rendered at $D(\cdot)$5 pixels and paired with floating-point depth maps, dense optical flow, camera intrinsics, ground-truth poses, and the full 3D mesh (Lena et al., 9 Feb 2026).

Benchmarking used Lite-Mono with $D(\cdot)$6 inputs, a ResNet-18 pose encoder, AdamW, learning rate $D(\cdot)$7, and 50 epochs. On the C3VD test subset, the C3VD-trained model achieved $D(\cdot)$8, $D(\cdot)$9 mm, and $1024 \times 768$0, whereas the RealSynCol-trained model achieved $1024 \times 768$1, $1024 \times 768$2 mm, and $1024 \times 768$3 (Lena et al., 9 Feb 2026). A recent transformer-based depth foundation model, DAM v2-small, required median scaling in zero-shot mode and yielded $1024 \times 768$4 and $1024 \times 768$5 mm; after LoRA fine-tuning on RealSynCol with rank 8, it achieved $1024 \times 768$6 and $1024 \times 768$7 mm without test-time scaling (Lena et al., 9 Feb 2026). This suggests that realistic synthetic supervision is now a central component of colonoscopy reconstruction research rather than a mere pretraining convenience.

5. Full-procedure video annotation and multimodal lesion understanding

Colon-Bench addresses the absence of densely annotated full-procedure colonoscopy video datasets by introducing 528 human-verified lesion windows, spanning 12.89 h and 464,035 frames, drawn from 59 patient sequences (Hamdi et al., 26 Mar 2026). The dataset includes 300,132 bounding boxes, 213,067 segmentation masks, and 133,289 words of clinical descriptions across 14 lesion categories: sessile polyp, pedunculated polyp, sessile serrated lesion, flat elevated lesion (Paris IIa), depressed lesion (Paris IIc), ulcer, erosion, diverticulum, angioectasia or angiodysplasia, bleeding, lipoma, erythematous lesion, submucosal bulge, and other vascular anomalies (Hamdi et al., 26 Mar 2026).

Its annotation pipeline is explicitly agentic and multi-stage. A Gemini-2.5-flash-lite vision-LLM first proposed 1,325 candidate lesion windows from full-procedure videos; a Gemini-3-pro verification agent pruned 422 windows; EdgeTAM propagated bounding boxes across frames; Gemini-3-flash removed another 306 false positives with box overlays; and a web-based surgeon review rejected 69 windows, yielding the final 528 clips (Hamdi et al., 26 Mar 2026). The benchmark supports binary lesion classification, open-vocabulary video object segmentation, and video VQA, using IoU, Dice, mAP, accuracy, precision, recall, and $1024 \times 768$8 as evaluation metrics (Hamdi et al., 26 Mar 2026).

Results reported on Colon-Bench are notable for the relative strength of generalist multimodal LLMs. For OV-VOS, Gemini 3 Flash achieved $1024 \times 768$9 mIoU and $768 \times 768$0 Dice, whereas SAM-3 achieved $768 \times 768$1 mIoU and $768 \times 768$2 Dice (Hamdi et al., 26 Mar 2026). For binary classification, Gemini 3.1 Flash-Lite reached $768 \times 768$3 accuracy and $768 \times 768$4 $768 \times 768$5, while Endo-CLIP achieved $768 \times 768$6 $768 \times 768$7 (Hamdi et al., 26 Mar 2026). For detection, Gemini 3 Flash achieved $768 \times 768$8 $768 \times 768$9 and $\pm 25^\circ$0 mAP@50, while GPT-4o had $\pm 25^\circ$1 $\pm 25^\circ$2 and SAM-3 failed at below $\pm 25^\circ$3 (Hamdi et al., 26 Mar 2026). The study also introduced a training-free “colon-skill” prompting strategy, consisting of distilled morphological cues and error-avoidance rules, which improved zero-shot VQA by up to 9.7 percentage points across several MLLMs (Hamdi et al., 26 Mar 2026). A common assumption in medical AI is that domain-specific systems necessarily dominate generalist models; these results do not support that assumption for lesion localization and video reasoning in colonoscopy.

6. Molecular oncology and the distinct algebraic meaning of “colon”

Transcriptomic analysis provides a molecular view of colon cancer distinct from imaging and endoscopy. In one integrated bioinformatics study, 98 tumor and 98 healthy colonic mucosa samples from GSE44076 were compared after normalization, $\pm 25^\circ$4 transformation, limma-based differential testing in GEO2R, and Benjamini–Hochberg correction (Behrouzifar, 2023). Under-expression was defined by adjusted $\pm 25^\circ$5-value $\pm 25^\circ$6 and $\pm 25^\circ$7, yielding 635 under-expressed genes in colon cancer (Behrouzifar, 2023). Protein–protein interactions retrieved via STRING and analyzed in Cytoscape identified 12 hub under-expressed genes with degree $\pm 25^\circ$8: CLCA1, SLC26A3, SI, KIT, HPGDS, NR1H4, ADIPOQ, PPARGC1A, GCG, MS4A12, GUCA2A, and FABP1 (Behrouzifar, 2023).

A three-way Venn comparison across colon cancer, extensive ulcerative colitis, and limited ulcerative colitis identified NR1H4 as the sole downregulated gene shared between limited ulcerative colitis and colon cancer (Behrouzifar, 2023). The study further associated CLCA1 with the “Pancreatic secretion” pathway and antagonism of Wnt/$\pm 25^\circ$9-catenin signaling, PPARGC1A with the “Adipocytokine signaling pathway” and suppression of NF-$1.0$0B-mediated inflammation, and AQP8 with “Bile secretion,” epithelial hydration, and barrier function (Behrouzifar, 2023). The interpretation advanced there is that loss of these regulators disrupts epithelial homeostasis and contributes to invasion, metastasis, and the inflammatory bridge between ulcerative colitis and neoplasia (Behrouzifar, 2023).

In mathematics, by contrast, “colon” denotes an ideal quotient rather than an anatomical structure. For ideals $1.0$1 over a field of characteristic zero, the colon ideal is

$1.0$2

and in particular $1.0$3 consists of all polynomials $1.0$4 such that $1.0$5 (Booth et al., 12 Mar 2026). For integers $1.0$6 and $1.0$7, the ideal is generated by two homogeneous elements whose degrees add up to $1.0$8 (Booth et al., 12 Mar 2026). When $1.0$9, the generators are $y' \in \mathbb{R}^3$00 and a single-sum form $y' \in \mathbb{R}^3$01; when $y' \in \mathbb{R}^3$02, explicit two-summand generators $y' \in \mathbb{R}^3$03 or $y' \in \mathbb{R}^3$04 arise according to the parity of $y' \in \mathbb{R}^3$05 (Booth et al., 12 Mar 2026). These formulas are then used to construct a determinant whose vanishing characterizes failure of the weak Lefschetz property for certain monomial almost complete intersections in three variables (Booth et al., 12 Mar 2026).

Across these literatures, the colon appears as a paradigmatic object of contemporary quantitative research: a deformable anatomical tube requiring patient-specific geometry in navigation, a histological substrate for cancer diagnosis, a segmentation target in CT and colonoscopy, a source of multimodal reasoning challenges for MLLMs, a locus of transcriptomic dysregulation in malignancy and inflammatory disease, and, in algebra, the namesake of a precisely structured ideal quotient.