Toward real-time quantification of fluorescence molecular probes using target/background ratio for guiding biopsy and endoscopic therapy of esophageal neoplasia

Abstract

Multimodal endoscopy using fluorescence molecular probes is a promising method of surveying the entire esophagus to detect cancer progression. Using the fluorescence ratio of a target compared to a surrounding background, a quantitative value is diagnostic for progression from Barrett's esophagus to high-grade dysplasia (HGD) and esophageal adenocarcinoma (EAC). However, current quantification of fluorescent images is done only after the endoscopic procedure. We developed a Chan-Vese-based algorithm to segment fluorescence targets, and subsequent morphological operations to generate background, thus calculating target/background (T/B) ratios, potentially to provide real-time guidance for biopsy and endoscopic therapy. With an initial processing speed of 2 fps and by calculating the T/B ratio for each frame, our method provides quasireal-time quantification of the molecular probe labeling to the endoscopist. Furthermore, an automatic computer-aided diagnosis algorithm can be applied to the recorded endoscopic video, and the overall T/B ratio is calculated for each patient. The receiver operating characteristic curve was employed to determine the threshold for classification of HGD/EAC using leave-one-out cross-validation. With 92% sensitivity and 75% specificity to classify HGD/EAC, our automatic algorithm shows promising results for a surveillance procedure to help manage esophageal cancer and other cancers inspected by endoscopy.

Keywords: computer-aided diagnosis; esophageal cancer detection; fluorescence video processing; multimodal endoscopy; surgical guidance; target/background ratio.

Fig. 1

T/B ratio calculation procedure for fluorescence image.

Fig. 2

Histogram-based method to remove artifact frames. A frame with T/B ratio equal to 1.0 was first removed from the histogram, and then a combination of statistics and histogram cutoff is applied for computing a single T/B ratio for the patient. In this example, this single value is calculated as mean T/B ratio of lower 90% of all frames.

Fig. 3

(a) Raw fluorescence frame, (b) image after $5\times 5$ kernel Gaussian filter, (c) segmentation result with parameter $\mu =1$, $\nu =60$, ${\lambda }_1=2.0$, ${\lambda }_2=2.1$, $dt=0.5$, $h=1$, (d) image after erosion applied to remove small isolated pixels, (e) image after dilation to generate surrounding background band with a width of 30 pixels (gray band), and (f) highlighted fluorescence target ROI (inside red contour) with T/B ratio greater than the set threshold.

Fig. 4

(a) Fluorescence image collected without processing, (b) T/B ratio 1.48, calculated with parameter $\mu =1$, $\nu =60$, ${\lambda }_1=2.0$, ${\lambda }_2=2.1$, $dt=0.5$, $h=1$, and $B=30\text{pixels}$, (c) T/B ratio 1.52, calculated with parameter $\mu =1$, $\nu =180$, ${\lambda }_1=2.0$, ${\lambda }_2=2.1$, $dt=0.5$, $h=1$, and $B=30\text{pixels}$. (d) T/B ratio 1.64, calculated with parameter $\mu =1$, $\nu =330$, ${\lambda }_1=2.0$, ${\lambda }_2=2.1$, $dt=0.5$, $h=1$, and $B=30\text{pixels}$. Comparing (b) to (d), with other parameters unchanged, more constraints were applied on the area size inside the contour, which resulted in smaller target area and increased T/B ratio.

Fig. 5

(a) Fluorescence image collected without processing, (b) T/B ratio calculated using surrounding background with a width of (b) $B=20\text{pixels}$, $\mathrm{T}/\mathrm{B}=1.57$, (c) 30 pixels, $\mathrm{T}/\mathrm{B}=1.64$, and (d) 40 pixels, $\mathrm{T}/\mathrm{B}=1.71$. Comparing (b) to (d), with other parameters unchanged, wider surrounding background resulted in higher T/B ratio.

Fig. 6

Example of T/B ratios calculated from continuous video frames (Video 1, MP4, 2.14 MB [URL: http://dx.doi.org/10.1117/1.JMI.4.2.024502.1]).

Fig. 7

AUC from different statistics and histogram cutoffs to calculate T/B ratio for patients. The best combination to compute T/B ratio for each video is median of lower 95% frame T/B ratios after removing frames with T/B ratio equivalent to 1.0, with AUC 0.875.

Fig. 8

(a) T/B ratios for individual patients grouped by the most advanced grade found on pathology. (b)The mean and standard deviation values for T/B ratio in each category. The mean values were 1.44 for SQ ($n=2$), 1.27 for BE ($n=3$), 1.26 for GEJ ($n=4$), 1.35 for LGD ($n=7$), 1.57 for HGD ($n=21$), and 1.56 for EAC ($n=13$). (c) At a target/background ratio of 1.33, sensitivity was 94% and specificity was 75%. At a target/background ratio of 1.45, specificity was 88%, and sensitivity was 66%. These are the results calculated using T/B ratios for all 50 patients. The final results will be calculated as the average from LOOCV (see above). (d) The ROC curve comparing the T/B ratios for early neoplasia (HGD and EAC) with those for the other classifications (SQ, BE, LGD, and GEJ) had an AUC of 0.875.

Fig. 9

Two methods to highlight the targets: (a) enhancing the color of the targets and (b) encircling the targets.