PaperlayerView exact revision on arXiv

Source-linked AI summary

Automated polyp detection in colon capsule endoscopy

Alexander V. Mamonov, Isabel N. Figueiredo, Pedro N. Figueiredo, Yen-Hsi Richard Tsai

arXiv:1305.1912v4cs.CV

TL;DR

Manual review of thousands of CCE frames makes automated polyp detection valuable for reducing operator burden. The paper combines texture pre-selection with geometric frame classification based on segmented protrusions and evaluates it statistically on five patients. It reports 47.4% per-frame sensitivity and 81.25% per-polyp sensitivity at approximately 90% specificity.

  • Problem

    Thousands of frames in CCE videos make manual analysis burdensome, motivating accurate automated polyp detection with a low false-positive rate.

  • Method

    The algorithm pre-selects frames by texture content and classifies remaining frames using geometric protrusion analysis, including a best fit ball radius and shape criteria.

  • Results

    81.25% sensitivity per polyp and 47.4% sensitivity per frame were obtained at approximately 90% specificity on the evaluated data.

  • Takeaways & Limitations

    The reported performance reduces the number of video frames requiring manual inspection and represents progress toward fully automated polyp detection.

Abstract

from arXiv · show

Colorectal polyps are important precursors to colon cancer, a major health problem. Colon capsule endoscopy (CCE) is a safe and minimally invasive examination procedure, in which the images of the intestine are obtained via digital cameras on board of a small capsule ingested by a patient. The video sequence is then analyzed for the presence of polyps. We propose an algorithm that relieves the labor of a human operator analyzing the frames in the video sequence. The algorithm acts as a binary classifier, which labels the frame as either containing polyps or not, based on the geometrical analysis and the texture content of the frame. The geometrical analysis is based on a segmentation of an image with the help of a mid-pass filter. The features extracted by the segmentation procedure are classified according to an assumption that the polyps are characterized as protrusions that are mostly round in shape. Thus, we use a best fit ball radius as a decision parameter of a binary classifier. We present a statistical study of the performance of our approach on a data set containing over 18,900 frames from the endoscopic video sequences of five adult patients. The algorithm demonstrates a solid performance, achieving 47% sensitivity per frame and over 81% sensitivity per polyp at a specificity level of 90%. On average, with a video sequence length of 3747 frames, only 367 false positive frames need to be inspected by a human operator.

I. INTRODUCTION

CCE offers a minimally invasive way to acquire extensive intestinal video, but thousands of frames make manual polyp review burdensome. The paper proposes automated frame classification using texture and geometric information, with parameters currently selected manually.

  • Motivation: Colorectal polyps are important precursors to colon cancer, motivating their detection during minimally invasive CCE examinations.CCE records intestinal images using a camera-equipped ingested capsule.
  • Motivation: Thousands of frames in a patient video make complete manual analysis burdensome, creating a need for accurate detection with few false positives.High sensitivity reduces missed polyps, while high specificity limits frames requiring human inspection.
  • Approach: The proposed algorithm classifies each CCE frame as containing polyps or normal tissue using texture and geometric information.Geometric processing treats polyps as protrusions and uses the radius of a best fit ball as a decision parameter.
  • Approach: Texture pre-selection discards frames with either too little or too much texture before binary geometric classification.Low texture is associated with flat mucosa, while excessive texture may indicate bubbles or trash liquids and increase false positives.
  • Scope: The algorithm requires numerical parameters chosen in advance, and the authors lack a systematic calibration procedure.Parameter values used in experiments were selected manually, leaving calibration as future research.

B. Texture computation and convolution

The algorithm measures frame texture through a texture–cartoon decomposition, nonlinear Gaussian convolution, and a scalar texture-content statistic. Lower and upper thresholds retain textured polyp candidates while rejecting flat mucosa and frames dominated by bubbles or trash.

  • Pipeline overview: Figure 1 traces processing from color input through grayscale normalization, texture transformation, filtering, segmentation, and feature-shape analysis.The caption identifies four connected components in the illustrated segmentation.
  • Texture decomposition: Texture computation begins by decomposing the pre-processed frame into texture t and cartoon c components.The decomposition uses an iterative low-pass Gaussian-filtering algorithm.
  • Convolution transform: The nonlinear transform emphasizes large textured regions while reducing the influence of small regions with strong texture.It applies pixel-wise exponentiation and Gaussian convolution; p < 1 de-emphasizes small strongly textured regions.
  • Pre-selection: The texture-content statistic Tmax is used in a threshold-based pre-selection that discards frames outside an acceptable texture range.The lower and upper bounds target insufficient texture and excessive texture, respectively.
  • Statistical behavior: 90% of polyp frames pass pre-selection, compared with 47.84% of normal frames, using the parameters in Table I.The distributions of Tmax show normal-frame peaks at lower values and elevated normal-frame counts at high values.

C. Mid-pass filtering and segmentation

After texture pre-selection, mid-pass filtering isolates protrusions within expected size limits and thresholding converts them into connected candidate regions. The segmentation threshold is bounded to avoid unstable extremes.

  • Mid-pass filtering: Geometrical processing targets polyp-like protrusions while filtering out features that are too small or too large.Expected polyp dimensions determine the two Gaussian scales used by the mid-pass filter.
  • Mid-pass filtering: The filtering ratio emphasizes relative prominence and is invariant under scaling of the image.Using the positive part of the ratio highlights protrusions as large positive values.
  • Segmentation: Thresholding the filtered image produces a binary segmentation whose connected components are processed as separate features.Each component is represented by binary pixel membership and indexed among NC components.
  • Segmentation: The segmentation threshold equals half the maximum filtered value unless that value falls outside prescribed lower and upper bounds.The bounds satisfy MU > ML > 0.

D. Geometrical processing and the tensor of intertia

Candidate regions are filtered by size and shape before a geometric decision is made. Tensor-of-inertia ellipses quantify elongation, reflecting the assumption that polyps are rounder than mucosal folds.

  • Feature selection: Feature-size filtering removes regions that are too large for likely polyps or too small to distinguish from filtering and segmentation artifacts.Large features typically correspond to normal mucosal folds.
  • Shape analysis: Tensors of inertia provide centers of mass and ellipses whose eccentricities measure how strongly each candidate feature is stretched.The ellipses are scaled so their areas match the corresponding feature sizes.
  • Shape analysis: Polyps are expected to be more round than mucosal folds, so the algorithm selects moderately stretched features using an eccentricity threshold.Figure 4 contrasts a polyp frame with normal frames containing pronounced folds.
  • Figure guide: Figure 4 compares original frames, mid-pass-filtered images, and binary segmentations for one polyp frame and two normal frames.The comparison supports evaluating whether candidate regions are compact or stretched.
  • Classification: The combined geometric criterion determines whether candidate features continue to the final frame-level decision stage.Frames without any feature satisfying the criterion are labeled normal.

E. Decision parameter and binary classifier

The classifier uses the maximum best-fit ball radius Rmax over geometrically suitable segmented features, with texture pre-selection and a threshold RP determining the frame label. The radius captures rounded protrusion size and complements two-dimensional segmentation by fitting the filtered image itself.

  • Decision parameter: The fit-ball radius is computed on the mid-pass filtered image u rather than binary segmentation s to incorporate protrusion height.The segmentation supplies feature geometry, while fitting to u retains intensity information about the protrusion.
  • Decision parameter: Rmax is the maximum optimal fit-ball radius among features satisfying the combined geometric criterion.For each qualifying feature, the radius is found by a one-dimensional search minimizing the Frobenius-norm fitting error.
  • Decision parameter: Frames failing texture pre-selection or geometric qualification receive Rmax = 0 before classification.This includes frames with no qualifying feature and frames whose texture maximum lies outside the accepted interval.
  • Binary classifier: The binary classifier labels a frame “polyp” when Rmax ≥ RP and “normal” when Rmax < RP.The discrimination threshold RP is selected through statistical performance analysis.
  • Observed classifications: Correct classifications include varied polyp shapes and cases with small amounts of trash liquid or nearby mucosal folds.Incorrect classifications arise from stretched features, subthreshold radii, insufficient illumination correction, and mucosal folds resembling polyps.
  • Statistical behavior: The radius distributions for polyp frames are shifted toward larger values than those for normal frames after pre-selection and geometric filtering.The normal-frame distribution remains lower partly because rejected frames are assigned Rmax = 0.

F. Summary of the algorithm

Algorithm 1 processes each capsule-endoscopy frame through preprocessing, texture-based pre-selection, segmentation, geometric analysis, and best-fit-radius classification. Its computational cost is O(NxNy), with the only minimization reduced to a simple one-dimensional search.

  • Pre-processing: Each input frame is converted to grayscale, intensity-normalized, and preprocessed outside the circular capsule mask.The resulting frame is passed to texture analysis.
  • Texture analysis: Texture analysis computes a transformed texture measure and discards frames whose maximum fails the pre-selection criterion.Discarded frames are assigned Rmax = 0 and labeled normal at the final step.
  • Geometric analysis: The filtered frame is segmented into connected components, which are screened using eccentricity and feature-size criteria.Only features satisfying the combined geometric criterion proceed to radius estimation.
  • Classification: For surviving features, the algorithm computes best-fit radii, takes their maximum Rmax, and applies the binary threshold to produce “normal” or “polyp.”The maximum is taken over the qualifying feature set.
  • Computational cost: O(NxNy) operations are required per frame, and the radius minimization is only a simple one-dimensional search.The reported implementation avoids expensive PDE or optimization problems.

IV. TESTING METHODOLOGY AND DATA SET

The study evaluates Algorithm 1 statistically on testing data, using Matlab and the Image Processing Toolbox for implementation and computation.

  • Testing methodology: The evaluation uses a statistical performance test of Algorithm 1 on a designated testing data set.The implementation and all computations were performed with Matlab and the Image Processing Toolbox.

A. Data set

The data set contains full-exam CCE videos from five adult patients, with many realistic normal frames and polyp sequences organized for both frame-level and polyp-level evaluation.

  • Data set rationale: A sufficiently rich data set is used because small samples can cause overtuning and misleadingly good performance.The stated concern is that an algorithm appearing successful on limited samples may fail under realistic conditions.
  • Data set composition: 18,968 frames comprise 18,738 normal frames and 230 polyp frames from five adult patients.The videos were captured with PillCam COLON 2 at 512 × 512 pixels before downsampling.
  • Data set composition: Normal frames include mucosal folds, diverticula, bubbles, and trash liquids to represent realistic examination conditions.The large number of normal frames supports evaluation of both sensitivity and specificity.
  • Evaluation units: The 230 polyp frames contain 16 polyps organized into sequences corresponding to individual polyps.This organization enables performance analysis per frame and per polyp.

B. Receiver operating characteristic curve

The study evaluates the binary classifier statistically using ROC curves, balancing sensitivity against specificity at varying discrimination thresholds. Because polyps appear across consecutive frames, performance is also assessed per polyp rather than only per frame.

  • Per-frame evaluation: Over 18,900 frames support a statistical evaluation using receiver operating characteristic curves rather than specially selected examples.The ROC analysis captures classifier performance as the discrimination threshold changes.
  • ROC analysis: ROC curves represent the trade-off between false-positive rate and true-positive rate as the discrimination threshold varies.Sensitivity is the true-positive rate, while specificity measures correct classification of non-polyp frames.
  • Threshold selection: The threshold is selected on a training subset by choosing the minimum discrimination value that achieves a desired specificity.Patient 4, with over 8,500 frames, is used as the training subset in the described procedure.
  • Per-polyp evaluation: Per-polyp sensitivity counts a polyp as detected when at least one frame in its corresponding sequence is classified as “polyp”.This complements per-frame sensitivity with a measure closer to the clinical objective of finding actual polyps.

C. The choice of parameters and robustness

The algorithm uses manually chosen geometric parameters, and the paper evaluates robustness by perturbing parameters individually and measuring resulting changes in specificity and sensitivity. The robustness study uses a reduced data set to limit computational effort.

  • Parameter choice: Algorithm 1 uses numerical parameters primarily describing the size or shape of geometric features, chosen manually from expectations about polyp morphology.The paper states that the parameter values were not fine-tuned for the particular data set.
  • Robustness measure: Robustness is assessed through relative sensitivities of specificity and sensitivity, computed separately for per-frame and per-polyp performance.Each sensitivity measures the change in a statistical quantity after perturbing one parameter.
  • Robustness measure: Each robustness calculation compares a base parameter value with a perturbed value while keeping all other parameters fixed.The paper defines the comparison using SPEC_base and SPEC_pert, and analogously for sensitivity.
  • Perturbation design: Parameters are perturbed to 1.1 · X_base, with integer-valued parameters rounded up, to capture nonlinear algorithm behavior.The perturbation is applied one parameter at a time.
  • Computational setup: The reduced robustness data set retains all polyp frames but limits normal frames to 4000 and restricts the evaluated parameters.The restriction emphasizes the geometric portion of the algorithm over pre-selection.

V. TESTING RESULTS

Testing on the five-patient data set shows that the classifier reaches roughly 90% specificity with moderate per-frame sensitivity and substantially higher per-polyp sensitivity. The evaluation also measures the resulting manual-review burden.

  • Evaluation design: The evaluation applies the statistical testing methodology to Algorithm 1, including per-frame and per-polyp analyses.The reported results include ROC studies and patient-level inspection burden measurements.
  • Per-frame results: 90.2% specificity and 47.4% sensitivity are obtained on the whole data set with threshold R_P = 37.The threshold was selected to meet a target specificity of 90% on the training subset.
  • Per-polyp results: 81.25% per-polyp sensitivity detects 13 of 16 polyps in at least one frame of each corresponding sequence at R_P = 37.Per-polyp performance is higher because each polyp can be represented across multiple consecutive frames.
  • Per-polyp results: 93.47% specificity is achieved while retaining 81.25% per-polyp sensitivity when the threshold is increased to R_P = 40.This result is reported for the per-polyp ROC analysis.
  • Manual-review burden: The study reports a massive reduction in frames requiring manual inspection across all five patients.False positives and false-positive rates per patient are used to quantify the remaining inspection burden.

B. Robustness

The algorithm is generally robust to parameter changes, with a notable exception for σ2, while broader limitations and future improvements remain.

  • Parameter robustness: Around 5% or below per-frame SENS changes occur for all parameters except σ2.The per-frame sensitivity is more affected than SPEC but remains limited for most parameters.
  • Parameter robustness: A 10% parameter perturbation does not affect per-polyp SENS except for the same σ2 case.Per-polyp sensitivity is reported as the most robust measure in the sensitivity study.
  • Parameter robustness: For σ2, per-frame SENS changes from 47.4% to 53.5%, while per-polyp SENS changes from 81.2% to 93.8%.σ2 affects the maximal feature size Rmax, which the algorithm uses to identify polyps.
  • Limitations and future work: The algorithm remains limited by frame-level localization, grayscale processing, manually chosen parameters, and reliance on a simple binary classifier.The paper identifies motion-related localization, omitted color information, calibration, and more advanced classifiers as improvement areas.
Loading 1305.1912v4…