Detection of Necrotizing Enterocolitis in Newborns ...

Detection of Necrotizing Enterocolitis in Newborns using Abdominal Thermal Signature Analysis G.M. Kouamou Ntonfo Dep. of Sc. and Methods for Eng. U. of Modena and Reggio Emilia Reggio Emilia,Italy [email protected]

Abstract— In this paper we present a novel approach to early diagnosis of Necrotizing Entercolitis in premature newborns. In particular, using an infrared thermal camera, thermal image of newborn abdomen is acquired. Image processing and spatial segmentation are then used to retrieve thermal signature which is represented by a sample distribution of values from an 8-bit grey level color palette. First order statistical features are then extracted from thermal signature and are used into a classifier. Preliminary results are encouraging and show the potential use of the proposed approach for classification between healthy and sick newborns.

Keywords— Necrotizing premature newborns

enterocolitis,

Erika Bariciak

Monique Frize Systems and Computer Engineering Carleton University Ottawa, ON, Canada [email protected]

thermal

image,

I. INTRODUCTION Necrotizing enterocolitis (NEC) is a serious neonatal gastrointestinal complication [1, 2], that is complex to diagnose, and occurs through the interaction of many physiological factors. NEC mainly occurs in up to 7% of infants born less than 1500g [3, 4] but it can also occur in term babies. It presents with a wide range of symptoms which are not specific to NEC and there can be a rapid deterioration that can lead to septic shock and death. When a case of NEC is suspected, treatment includes cessation of feedings, provision of intravenous nutrition and antibiotics, and treatment for shock. Perforation or ongoing deterioration may require surgical intervention. The mortality rate is 15-30%, but postoperative mortality increases to 40% [1]. Survivors may suffer significant complications, including long-term impaired bowel function, and are at an increased risk of developing severe neuro-developmental disabilities [5]. The condition involves massive inflammation and peritoneal reaction which leads to significant pain and tenderness. Early detection can lead to timely intervention and decrease the overall morbidity, mortality, and hospital expenses [6, 7], however it remains a significant challenge to diagnose NEC at an early stage. Abdominal radiographic imaging and ultrasonography are the most common diagnostic modalities currently used [8, 9] but these tools do not reliably detect NEC in its early stages. In most institutions, the staging of NEC is largely based on the use of modified Bell’s criteria, with definitive NEC signified

Dep. of Paediatrics University of Ottawa Ottawa, ON, Canada [email protected]

by Bell’s stage 2 or higher [10]. The inflammation and pain associated with NEC can be present prior to the appearance of significant changes on x-ray images [11,12]. Infrared imaging (thermography) is a non-invasive tool that captures the thermal distribution of the human body [13] that can provide valuable information about the inflammatory and painful processes involved in NEC [14, 7] and may be able to help diagnose NEC earlier than more traditional imaging. Thermograms have been used to analyze temperature distributions between central and peripheral body in extremely low birth weight newborns and to examine the relationship between body temperature and development of NEC in premature babies [15]. In other work, statistical feature extraction, infrared temperature profiles of abdomen quadrants were submitted to a pattern classification tool (decision trees) to discriminate between infants with and without NEC [7]. In this new work, we developed a novel detection approach based on the segmentation of the abdominal area into identical small areas from which a thermal signature, represented by a vector of mean values of resulting areas, is extracted. First order statistical features of the thermal signature are then used, through a pattern classification approach, to identify inflammation associated with NEC in premature infants and to discriminate between healthy and NEC babies.

I. METHODOLOGY A. Image Acquisition Following approval of the Ethics Review Boards at the Children’s Hospital of Eastern Ontario (CHEO) and Ottawa Hospital General Campus (OGH), informed consent was obtained from parents of babies in both NICUs. Infrared images were acquired from infants who were between 23 and 32+6 weeks gestational age between Dec. 2006 and Jan. 2010. Healthy infants were chosen when they did not exhibit any clinical or radiographic signs associated with NEC and were not affected by any condition that alters thermal regulation. The NEC-affected babies exhibited clinical signs of NEC, with radiographic evidence of Bell’s stage 2 or higher. Babies were excluded if they had intra-abdominal congenital anomalies,

Fig. 3 (a) ROI selection.

Figure 1: Thermal image acquisition procedure (courtesy of CHEO)

active sepsis, or umbilical or abdominal dressings occluding the imaging area [7]. Premature infants were in a thermally controlled environment from which they were removed for 45-60s. Infrared images were collected using a long-wavelength IR camera (uncooled microbolometer focal plane array, 320 x 420 pixels, thermal and spatial sensitivity of 0.05°C (at 30°C) and 1.3 mrad, respectively). The IR camera was approximately 60cm from the baby’s abdomen facing down. Figure 1 shows a thermal image acquisition procedure [7]. Fig. 1 illustrates the image collection set-up.

Fig. 3 (b) Noisy ROI.

Axial and rectal temperatures were taken before and after

Fig. 3 (c) Denoised ROI. Figure 3: Abdominal thermal image processing steps.

Figure 2: An example of newborn abdominal thermal image

the procedure to record significant drops in temperature while outside the incubator and nurses assessed the baby’s level of pain during the imaging using the Premature infant Pain Profile (PIPP) Score, a standardized NICU pain scoring tool based on infant behaviour and vital signs [16, 7]. The raw images from the thermal camera are bounded to a temperature interval between 35°C and 38°C. Images are exported on a bitmap image file at 320x240 pixels resolution and 8 bit grey color palette, so that pixel’s interval value are between 0 and 255.

B. Image processing A thermal image, in grey scale color palette of a newborn abdomen, could be poor in contrast making it potentially difficult to distinguish temperature variations or patterns with the naked eye. Since we measure grey level values of pixels and those values are related to effective temperature values of the newborn body, we avoid any contrast enhancement procedure which aims to enhance human visual perception of temperature variation by distorting measured values, therefore introducing noise. A rectangular area over the umbilical stump is selected and extracted from the thermal image, Figure 3a, represents the Region Of Interest (ROI) where the thermal signature is analysed. The ROI has to be nearly the same for all newborns thermal images so that they can be compared.

process is then computed on the ROI area. The ROI’s denoised thermal image is shown in Figure 3c. On newborns, the dried umbilical stump has a completely different temperature characterization with respect to other areas of the abdomen; the minimum area surrounding the navel is set to zero and used to exclude the navel area from further analysis.

Figure 4: Thermal signature over ROI, navel region value is set to zero.

Registered images are noisy due to the image collection process and the thermal camera sensor’s inherent noise [17] as shown in Figure 3b. The denoising process has to be efficient without changing considerably the image information and texture. Noise is considered a Gaussian white noise and denoising is performed using Non-Local Means Denoising algorithm [18]. While the traditional denoising methods goal is to replace the color of a pixel with an average of the values of nearby pixels, in Non-Local Means Denoising algorithm, the new value of a pixel is computed by considering a large neighbourhood, taking the mean weighted by a function of similarity of each pixel with the target pixel. The denoising

C. Thermal signature The ROI is subdivided horizontally and vertically to obtain identical small areas. For each area, the mean value of the underlying pixels is computed, and the area value is set to the mean. Values of small areas that intersect the umbilical area are set to zero and are not be considered in the next steps. An example is shown in Figure 4. The vector of small area values, namely thermal signature, is a vector of grey level values corresponding to a sampling of temperature measurements on the whole abdomen obtained from each area. From the temperature distributions, first order statistical measures are computed: Mean, Median, Standard deviation (SD), Median absolute deviation (MAD), Interquartile range (IQR), Total sum of squares (TSS), Kurtosis, Skewness. In the presence of a large dataset, we can then use features previously calculated to differentiate between Normal and NEC babies with the help of a classifier. TABLE I.

Normal NEC

SHAPIRO-WILK NORMALIY TEST RESULTS.

Samples

W

P value

Decision

80 84

0.985 0.984

0.492 0.406

Normal at 0.05 level Normal at 0.05 level

TABLE II.

Mean SD Median MAD Skewness Kurtosis IQR Min Max TSS

STATISTICAL MEASURES Normal

NEC

166.03 21.08 165.47 12.89 0.005 0.11 25.12 112 213 35104

146.75 25.25 148.18 16.82 0.049 -0.68 35 97.17 199 52926

II. PERFORMANCE ANALYSIS

Figure 5: Thermal signature distribution

Preliminary results are based on the analysis of thermal images of the abdominal area of two newborns. The first newborn did not present any pathology; the second newborn developed stage 2 NEC. After image processing, the width and height of the ROI are divided by 10 to obtain 100 temperature measurements samples. Thermal signatures represented by

grey level values are then extracted, excluding the umbilical area. In the first newborn, a generated distribution is made-up of 80 elements while in the second newborn it is made of 84 elements; both are normally distributed as shown by the results of Shapiro-Wilk test in Table I. Statistical features of the grey level values distributions are computed and reported in Table II

NEC. In future work, using a larger population of newborns, all relevant features of computed first order statistical measures will be identified and an unsupervised learning algorithm will be used to classify potential cases of NEC.

REFERENCES [1] [2] [3] [4] [5]

[6]

[7]

Figure 6: Boxplots representation of grey values distribution.

[8]

and the corresponding boxplots are shown in Fig. 5. A lower mean and median in the NEC baby compared to the normal baby may suggest that in this case, the measures are influenced by intestinal areas affected by necrosis and poor blood flow, two processes known to be involved in more advanced stages of NEC. In fact, the NEC boxplot and skewness values show that the mass of the distribution is concentrated between the first quartile and the mean, while the Normal boxplot and skewness values show that samples grey values are equally distributed around the mean. In absence of any pathology, the abdominal temperature measurements are very similar; small variations are due to measurement errors and physiology variations, therefore the extracted grey values distribution is leptokurtic (positive kurtosis value), but in the presence of NEC, a major dissimilarity between temperature measurements and the mean yields a platykurtic (negative kurtosis value) grey values distribution as shown in Fig. 6. III. CONCLUSION In this study, we have presented a novel approach to early diagnosis of the presence of NEC through thermal image analysis. Image processing and spatial segmentation were used to extract the thermal signature from the thermal images of newborn abdomens. Thermal signature is represented by a sample distribution of values from an 8-bit grey level color palette. First order statistical features are extracted from the thermal signature. Preliminary results show that IQR and kurtosis measures are good discriminants for the detection of

[9]

[10]

[11]

[12]

[13]

[14]

[15]

[16]

[17] [18]

P. W. Lin and B. J. Stoll, “Necrotising enterocolitis,” The Lancet, vol. 368, no. 9543, pp. 1271–1283, 2006. A. M. Thompson and M. J. Bizzarro, “Necrotizing enterocolitis in newborns,” Drugs, vol. 68, no. 9, pp. 1227–1238, 2008. J. Neu and W. A. Walker, “Necrotizing enterocolitis,” New England Journal of Medicine, vol. 364, no. 3, pp. 255–264, 2011. M. C. Henry and R. L. Moss, “Necrotizing enterocolitis,” Annual review of medicine, vol. 60, pp. 111–124, 2009. K. L. Schnabl, J. E. Van Aerde, A. B. Thomson, and M. T. Clandinin, “Necrotizing enterocolitis: a multifactorial disease with no cure,” World journal of gastroenterology: WJG, vol. 14, no. 14, p. 2142, 2008. J. A. Bisquera, T. R. Cooper, and C. L. Berseth, “Impact of necrotizing enterocolitis on length of stay and hospital charges in very low birth weight infants,” Pediatrics, vol. 109, no. 3, pp. 423–428, 2002. R. Nur and M. Frize, “Image processing of infrared thermal images for the detection of necrotizing enterocolitis,” in SPIE Medical Imaging. International Society for Optics and Photonics, 2013, pp. 86 692M–86 692M. S. Oh, C. Young, N. Gravenstein, S. Islam, and J. Neu, “Monitoring technologies in the neonatal intensive care unit: implications for the detection of necrotizing enterocolitis,” Journal of Perinatology, vol. 30, no. 11, pp. 701–708, 2010. M. Epelman, A. Daneman, O. M. Navarro, I. Morag, A. M. Moore, J. H. Kim, R. Faingold, G. Taylor, and J. T. Gerstle, “Necrotizing enterocolitis: Review of state-of-the-art imaging findings with pathologic correlation 1,” Radiographics, vol. 27, no. 2, pp. 285–305, 2007. M. Walsh and R. Kliegman, “Necrotizing enterocolitis: treatment based on staging criteria.” Pediatric clinics of North America, vol. 33, no. 1, pp. 179–201, 1986. C. Young, R. Sharma, M. Handfield, V. Mai, and J. Neu, “Biomarkers for infants at risk for necrotizing enterocolitis: clues to prevention?” Pediatric research, vol. 65, pp. 91R–97R, 2009. B. L. Frost, T. Jilling, and M. S. Caplan, “The importance of proinflammatory signaling in neonatal necrotizing enterocolitis,” in Seminars in perinatology, vol. 32, no. 2. Elsevier, 2008, pp. 100–106. B. F. Jones, “A reappraisal of the use of infrared thermal image analysis in medicine,” Medical Imaging, IEEE Transactions on, vol. 17, no. 6, pp. 1019–1027, 1998. C. L. Herry and M. Frize, “Quantitative assessment of pain-related thermal dysfunction through clinical digital infrared thermal imaging,” BioMedical Engineering OnLine, vol. 3, no. 1, p. 19, 2004. R. B. Knobel, B. D. Guenther, and H. E. Rice, “Thermoregulation and thermography in neonatal physiology and disease,” Biological research for nursing, vol. 13, no. 3, pp. 274–282, 2011. B. Stevens, C. Johnston, P. Petryshen, and A. Taddio, “Premature infant pain profile: development and initial validation,” The Clinical journal of pain, vol. 12, no. 1, pp. 13–22, 1996. S. Kogan, Electronic noise and fluctuations in solids. Cambridge England New York, NY, USA: Cambridge University Press, 1996. A. Buades, B. Coll, and J.-M. Morel, “Non-Local Means Denoising,” Image Processing On Line, vol. 1, 2011.